Changelog

Every release since 1.0, newest first, with what was in it and where it was announced. Below that, what shipped before there were versions.

1.61, Direct xAI routing, 2026-09-28

Grok now runs on xAI's own API too. Pick the version marked (direct) under Grok 4.7, Grok 4.6 or five more, and it costs xAI's list price, shows it is thinking from the first second, has twice the room to think, and keeps the conversation cached between turns so a follow-up costs a quarter to read. The gateway versions stay where they were.

New features

• Grok (direct). Grok 4.7, Grok 4.6, Grok 4.5, Grok 4.3, Grok 4.20, Grok 4.20 Multi-Agent and Grok Build 0.1 straight from xAI, listed under each model as its (direct) version, in chats and in rooms.
• Room to think. A direct reply can run to 64,000 tokens, thinking included, and the chat shows the model thinking from the first second instead of a blank wait. Grok 4.7 and 4.6 take the four thinking depths; 4.5 the first three.
• Cached conversations. The transcript is cached between turns automatically, so a follow-up reads the history at a quarter of the input price; there is no charge for writing the cache.
• List price, to the tick. A direct version costs what xAI charges, and the receipt is xAI's own figure for the call, cache discount included, not an estimate from a rate card. On six of the seven that is 5.5% under the gateway's copy; on Grok 4.7 the gateway's copy is the cheaper one, and the picker lists it first.

Improvements

• Receipts name the source. A run on a direct version says (direct) on its receipt, and the ledger records where it ran.
• Pictures as pictures. A direct version reads a picture you attach as the picture. PDFs and other files go as their text, since xAI does not take a file.
• If xAI is busy, out of credit or unreachable, the turn runs on the gateway's version at its price, and the receipt names the one that ran. A refusal never does.

Bug fixes

Nothing this time.

Also

• Content rules. The refusal pause covers this lane on its own, as it does the other two: an account xAI refuses five times in a day loses the Grok (direct) versions for a day (W416, on the error codes page), and every other model keeps running.
• Privacy. The privacy notice and the docs privacy page name xAI behind the Grok versions marked (direct): it receives the request with a fixed code made from your account id, is asked not to store the conversation, keeps requests up to 30 days to check for abuse and does not train on them.
• API. The direct versions are not on the public API; its Grok models run through the gateway as before.

1.6, askr Pro, 2026-09-27

A model you never have to pick. A new chat opens on askr Pro, Balanced: it reads each message with the room's memory and the tier, picks the door and the model, starts a skill when a line calls for one, makes the pictures, clips and sounds the line asks for beside the words, packages the lot around what you asked with a preview beside the room, and shows every pick under your line with Change.

New features

• askr Pro in the picker. The first row of the model picker in a chat or a room, with Cheap, Balanced and Performance beside it. Choose it and askr Pro picks the door and the model for every line at that tier, and the picker's trigger shows the mark and the tier word. The tier is a ceiling and a floor: Cheap never runs a dear model, Performance always picks the best there is, and a hard question moves up to the tier's reasoning models, never out of the tier, whatever the message says.
• On from the start. A new chat opens on askr Pro, Balanced, and so does the Ask composer on the home page, whose first send makes a room on it. Pick a model or a skill by hand and that sticks; the tier, or the model you picked instead, is remembered on that device. In a room the creator's pick sets askr Pro for the room and anyone else's holds for their own lines.
• What it reads. The message and what is attached to it, the last lines of the thread, the Building and Made lines of the room's memory, the tier, and the doors the room has on. One call to a decision model that answers with probabilities, not words, says how likely each door is, which of thirteen kinds of task the line is, whether it wants the web, how hard it is, whether it carries on the last one and whether it also wants something made; plain code turns the numbers into a door and a model, so nothing you type can talk it out of the tier. The decision has two and a half seconds and runs before the credits are held; when it is slow or down the line still runs, on the room's model through Ask, and the pick says so.
• Every pick shown. Under every line askr Pro decided, one line says what ran and why: askr Pro: Claude Sonnet 5 for writing, Balanced; askr Pro kept Claude Sonnet 5, this continues the last answer; askr Pro: Look things up skill on Claude Sonnet 5, Balanced. Change beside it opens the picker; pick a model by hand and the line runs again on it, askr Pro is off for the room from then on, and the model you picked is the room's model. While it works the room says askr Pro is choosing, then which model it picked, and a clip or a sound names the renderer or the voice on its wait row.
• No doors to choose. With askr Pro on the door chips under the message box are gone: it picks Ask, Code, Image, Video or Audio for every line itself, and there is no estimate beside Send, since the price is not known before the pick. A model picked by hand brings the chips and the estimate back. A line that opens by naming a picture, a photo of, a portrait of, a logo for, always paints, and a read-aloud ask in any of the common languages is read aloud.
• Made beside the words. When a line wants a picture, a clip or a sound as well as words or code, a site and its branding, a plan and its poster, the answering model makes them itself as it writes: it writes each prompt, each one runs through the room's own door on the tier's painter, renderer or voice, every made thing lands under the answer as its own line, and making a picture shows while it does. Nothing sends you to a door to do it yourself.
• Packaged around the ask. Code carries the pictures it was made with, by their addresses, instead of a file name you would have to supply. After a run that made things askr Pro checks once, off the turn, whether everything asked for is there and connected; whatever is missing is made in one more run, and a closing card under the last reply sums the package up, Your dog grooming site: 2 pages, logo, hero, gallery, 2 promo clips, 3 blog posts, with the bill for the whole line in credits and about how many dollars, split by clips, pictures, sounds and code or words, a clip still rendering counted when it lands.
• The preview takes the room. When a thread has a page to show, on a desktop the rail folds away and the panel takes the right half at full height, Preview first, with a small button that reaches Room settings and People; on a phone one switch moves between the chat and the full-screen preview, with Chat to come back. The page is assembled from the answer's files, its stylesheets and scripts inlined by name, even when the fences carry no names, and every picture, clip and sound made in the thread is put into the page before it is shown. Anchors scroll, a link to another page of the answer shows that page with Back, a link to a made picture, clip or sound opens it in the room, an outside link opens a new tab, and the frame never navigates away.
• Files, and Made here. A Files tab beside Preview and Code lists everything the thread made, pictures, clips, sounds and code files, each with Open and Save, a rendering clip saying so. A Made here card in the rail, and at the top of the Room sheet on a phone, shows the newest six pictures and clips, solo or collab, with All opening the Library.
• Code is not a wall. In every answer, in rooms, chats and skills, the fences fold into one row, Your code, 3 files, closed until you open it; the files are listed by name with Copy and Save and open one at a time. A short snippet with no file name stays inline.
• Skills, without asking. A line that calls for one of the ten skills, looking something up, writing, learning, fixing something, health, translating, money, ideas, a summary or shopping, runs with that skill's instructions and screen in the same thread: the sources, steps, cards or calculators arrive under the answer, and the pick says which skill ran. A thread that already carries a skill keeps its own.
• A follow-up keeps the model. When a line carries on the last task, the model that answered last answers again, as long as it is still live and inside the tier; the pick says kept. A new task picks afresh.
• Asked before a dear door. A picture or a sound needs a confident read, a clip a more confident one, and a follow-up in a thread whose last line came through that door a lighter one. Under the bar the line goes through Ask with a chip under it, Make it a picture instead, or a clip, or a sound, that runs the same line through that door. Ask and Code need no bar.
• Never stopped in your tracks. When a supplier declines a picture, a clip or a sound under its content rules, askr Pro rewords the request once, the brand, company, real person, franchise or copyrighted work replaced by the kind of thing it is, says so in the room, and tries again. A second refusal becomes an offer on your line, what could not be done and what would be done instead, with a chip, Yes, make the sound that way, that runs it. Inside a written answer the model rewords and tries again itself. With askr Pro off a refusal reads as before.

Improvements

• Search the web, when a line wants it. When the decision says a line wants the web, Search the web is on for that line, under the room's usual rule.
• A painter per line in rooms. The Image door in a room takes a painter picked for that one line, from the composer or from askr Pro, without changing the room's model.
• A file on the line is read. A line with a document or a data file attached that reads as an unsure Code pick runs through Ask instead, where the model reads the file.
• The tier words. The first tier reads Cheap, in the picker, the skill tier pill and the docs.
• The picker on a phone. The trigger shows the mark and the tier word alone, so it fits the narrow pill.
• Admin. An askr Pro page in the admin area: decisions in the last day and week, fallbacks, kept, changed by hand, media doors picked, average and p95 time, cost, by door, tier and class, the confidence buckets, rooms on askr Pro by tier, and the newest decisions.

Bug fixes

• In a room, picking a painter for one line answered W201; the line now runs on the painter it named.
• A picture call has four times the output room it had, and a picture cut at the budget (W513, four times in a row in one room) is asked for once more on the same hold, so the check after the run no longer has to make them all again.

Also

• Docs. askr Pro has a page: what it reads before each pick, how it decides, the rules, what it makes beside the words and how it packages it, what happens when a supplier refuses, what it costs and where it is not offered; askr Skills says a skill can start on its own and askr Collab names the picker's first row.
• What it costs. Each decision is a fraction of a credit, metered under its own source and charged at zero while askr is feeless, and so is the rewording of a refused request. The model that answers, and every picture, clip and sound it makes, is priced as always; nothing asks before a dear pick, the credits show under the answer as on every answer, and a line that made things shows the whole bill on its closing card. A line that asks for a site, a clip, a song and four pictures at Performance can run to thousands of credits, and the card says what it came to.
• Where it is not. A skill thread you opened yourself keeps its tier pill, since the tier there already picks the model; Compare names its own models and never takes askr Pro; the API names a model on every request and does not take it.
• Privacy. The decision reads your line the way any model you run does, and the record askr keeps of each decision, for the admin page and for version two, carries no message text.
• Measured before it shipped. Over 260 test lines, in six languages, with files, follow-ups and lines that try to name a dearer model, the door was right 98 to 99 times in 100 at every tier, the web read 97 to 98, and a decision took about 280 ms in the middle and under 400 at the 95th, for a fraction of a cent. A room driven through every door, the tier switch, the chips, Change, askr Pro off and the refusals passed 49 checks of 49.
• What is next. Version two builds the ladder around you, from what you reach for, ask and keep: the fusion step on the roadmap.

1.51, Direct OpenAI routing, 2026-09-27

GPT now runs on OpenAI's own API too. Pick the version marked (direct) under GPT-6 Astra Pro, GPT-6 Sol or five more, and it costs OpenAI's list price, shows it is thinking from the first second, has twice the room to think, and keeps the conversation cached between turns so a follow-up costs a fraction to read. The gateway versions stay where they were.

New features

• GPT (direct). GPT-6 Astra Pro, GPT-6 Astra, GPT-6 Sol, GPT-6 Luna, GPT-5.3-Codex, GPT-5.4 Mini and GPT-5.4 Nano straight from OpenAI, listed under each model as its (direct) version, in chats and in rooms.
• Room to think. A direct reply can run to 64,000 tokens, thinking included, and the chat shows the model thinking from the first second instead of a blank wait.
• Cached conversations. The transcript is cached between turns, so a follow-up within half an hour reads the history at a tenth of the input price. Writing the cache on the first turn costs a quarter more on GPT-6; on a long chat the reads take most of the input cost off.
• List price. A direct version costs what OpenAI charges, 5.5% under the gateway's copy on every token.

Improvements

• Receipts name the source. A run on a direct version says (direct) on its receipt, and the ledger records where it ran.
• PDFs as files. A direct version reads a PDF as the file itself, its text and its pages, and pictures as pictures.
• If OpenAI is busy, out of credit or unreachable, the turn runs on the gateway's version at its price, and the receipt names the one that ran. A refusal never does.

Bug fixes

Nothing this time.

Also

• Content rules. The refusal pause now covers each maker's direct versions on their own: an account OpenAI refuses five times in a day loses the GPT (direct) versions for a day (W416, on the error codes page), and Claude (direct) keeps running, and the other way round. Every other model keeps running.
• Privacy. The privacy notice and the docs privacy page name OpenAI behind the GPT versions marked (direct): it receives the request with a fixed code made from your account id, is asked not to store the conversation, keeps requests up to 30 days to check for abuse and does not train on them.
• API. The direct versions are not on the public API; its GPT models run through the gateway as before.

1.5, askr Mind, 2026-09-26

What you told one model is there for the next. Every chat and every room keeps a short memory that goes first into every model's prompt, with one more about you across chats, and all of it is yours to read, pin, edit, forget or switch off. A new chat is now a room of one, the skills and Compare work inside rooms, and a project can hold a team, so a room knows what the other rooms in it are doing.

New features

• A memory every model shares. After every answer askr Mind notes, in one line each, who is here, what is being built, what was decided, what is open, what was made and what each person prefers. The block goes first into every prompt, every model in a room reads the same one, and the chat says Using askr Mind while a model is being briefed. Nothing is rewritten and nothing is deleted on its own: entries are only added or replaced.
• Yours to keep or not. The askr Mind card in Room settings lists what the room remembers, click to edit, Forget on any line, Forget everything with a second press, and Remember this room for the creator. Pin beside Reply puts a line into the memory. After an answer a Mind noted chip says what was just remembered.
• Made, with captions, and Prompt used. Every picture, clip and sound made in a chat or a room is listed under Made; a picture gets a one-line caption of what is in it. When you ask a door for something, the prompt the generator gets is written from the memory and your request, and Prompt used under the result opens to the exact words, with the picture it built on. Ask for a made picture, moving, and it goes to the clip as the start frame.
• Recall. When a thread holds more than the prompt carries, the words of your request are searched across every line of the room and up to five earlier lines that match are handed to the model, each with who said it.
• About you, across chats. One list of who you are, what you build and how you like things, learned from your chats and put in front of every one of them. Settings, Memory shows it, with Remember me across chats, the entries, and Forget everything. A temporary chat, from the dashed bubble in the message box, keeps askr Mind out of that chat entirely.
• Projects. One memory shared by the rooms and chats you put in it: make one under Settings, Memory, pick it beside a chat's title or in Room settings, and every prompt in those rooms and chats carries it. Shared brings what askr knows about you into its rooms too; Walled off keeps you out.
• Your memory in a room, your choice. Bring my memory into this room, in Room settings, is each member's own switch for that room. A public room never takes it.
• Team projects. A project has people in it: invite by @ under Members on the Memory page, up to twelve, and anyone in it can put their private rooms in. The project's memory then says in one line per room what each room is doing, keeps every picture, clip and sound made in any of them with the room it came from, and a question in one room reaches the lines of the others: ask what is happening in the front-end room and the model names the room and the work. A picture asked for in one room can start from a picture made in another. A public room never joins, and leaving takes your rooms out with you.
• Recall by meaning. Every room line and memory entry is embedded, so a question about the thing we said about pricing finds it without the word pricing in it, beside the search by word. A long room stays reachable to its first line.
• Made from memory, with references. The prompt a picture, clip or sound door receives is written from what the room remembers, and the picture it builds on is handed to the run: a clip starts from the picture, an edit starts from the picture. Prompt used under the result shows the words, and a Prompt from what the room remembers preview appears under the message box while you type in a media door.

Improvements

• A room of one. A new chat is a room with one person in it, so a chat can do what a room does: threads, New thread from any line to continue from there, Re-run on the newest answer, a model per line from the composer, receipts that say the time taken and the pot that paid, and Export. It pays from your balance as a chat always did, free credits first, then given, then bought with the holder discount, and a run started alone settles alone even if you invite someone while it runs. Recent lists your rooms.
• Skills and Compare in rooms. A skill opens as a room whose thread carries the skill: the banner names it, Change starts a new thread under another skill, and the settings live on the thread so everyone in the room sees the same ones. Compare beside the model picker sends one line to two to four models side by side; every run holds its credits before any starts, Pick keeps one, and later lines and the memory read the picked answer only.
• Rooms say who spoke and what was made. A model in a room reads each earlier answer with the model that wrote it, and a picture, clip or sound as a line that says which door made it, with its caption. It is told which doors the room has and to say which one makes what you asked for, and never to pass a picture off as its own.
• Estimates that match receipts. The price beside Send in a room adds what the memory and a skill's brief put in front of the model, so the number before matches the receipt after.
• A room of one reads as a chat. No purse controls until someone is invited, tools and receipts in sentence case, a one-time note under the first answer saying the room remembers, and the project the room is in named on its title line.
• Continue in a room. An older chat gets Continue in a room, on the chat and in History: its lines and pictures become a room of one with the memory built at once.
• Every door starts a room. The Images, Video and Audio doors on the home page open a room set for that door on the first send, the start frame riding along.
• A memory that stays tidy. A near-duplicate of a line already remembered is refused, Who is here says what each person does, a Pick sets the thread's model, clips and sounds get a caption too, and a room folds at most once every twenty seconds however busy it is.
• Put this in a project? A new room whose plan resembles one of your projects offers to join it, once.
• Admin. An askr Mind page in the admin area: entries by scope and heading, the writer's calls and their cost, the rooms with the most memory.
• Handed on in rooms too. A Claude model that thinks through its whole budget in a room and never writes is handed to GPT-6 Sol under its own name, as the chat has done since 1.395, and nothing is charged for the first run.

Bug fixes

• A room of one short of credits for a picture or a clip now says so as a chat does, with the balance to top up, instead of the room wording.
• A sound made in a chat is described to the model as a sound; it used to be described as a picture.
• A picture asked for in a young room with the words make them, right after a message about three pictures, reached the image model as those bare words; the prompt writer now reads the last lines even before the room's first memory entry.
• An image model that answers a room in words instead of drawing is asked once more, plainly for a picture, on the same hold.
• Re-run on an answer ran the wrong line: it now runs the question the answer sits under.
• A wrapped caption was cut at its line break, and a writer that ran out of room kept a half-written last line; both now come through whole or not at all.

Also

• What it costs. askr Mind's own calls are metered under their own source and charged at zero credits while askr is feeless. They never hold a turn up, and a failure in askr Mind never fails a turn.
• Docs. askr Mind, askr Collab and askr Skills each have a page.
• Privacy. The privacy notice says what askr Mind keeps, that a temporary chat is neither read from nor written to, that your memory enters a room only when you switch it on there and never a public one, and that deleting a chat, a room or your account deletes its memory with it.

1.48, ASKR Connect improvements, 2026-09-27

Verify holdings is smoother for apps that buy through a router, as pump.fun does. When a buy like that cannot show whose wallet it is, the sixty seconds to sell start at once, and the sale you make is the proof.

New features

Nothing this time.

Improvements

• Straight to the sale. When a buy is turned down, the page shows the sixty seconds to sell right away, with Done, I sold. There is no button to press first, and a sale made after the verdict counts.
• The sale step is part of your check. Checks are limited to three an hour, and the sale step used to take one of them. It now belongs to the check you started, so it never uses one up.
• A contract says so at once. A sale from an address that is a contract cannot prove it either, so the page says that straight away and points you to Connect wallet, where signing does.

Bug fixes

Nothing this time.

1.47, Model audit, 2026-09-26

A user tested three of our models and said they were older than their labels. We tested all 383, published everything, and shipped what the test taught us.

New features

• The model audit report. Every model on askr against its label: the user's test, all 383 chat models with their verdicts, what we changed, and how to check for yourself.

Improvements

• Every model gets today's date. One line in front of every chat and room turn, so a model no longer calls last year's news "not yet happened". The API passes your messages unchanged.
• Cannot tell yet on the catalog. The 55 models the audit could not settle say so on their card, with the reason.
• Old names off the list. gpt-3.5-turbo-0613, gpt-3.5-turbo-16k and wizardlm-2-8x22b answer like newer models, so they are off until our supplier says what serves them. Eleven models that no longer answer are off too, and the API refuses a delisted id instead of running it.

Bug fixes

Nothing this time.

Also

• Refund. Everyone who ran Grok 4.7 got back what its route's hidden prompt tokens cost them, with no need to ask.

1.46, Rounding off, 2026-09-26

A tidy-up release. A public collab you share now brings you referrals, only one sound plays at a time, every error code explains itself when you tap it, and GIFs move in collab chats.

New features

• Public collabs refer. Share a public collab you made, and anyone who makes their askr account from it becomes your referral: 10% of what they top up comes to you in credits, the same as your own link. An account that already exists, a private room, or a sign-up that came with someone's code does not count. Referrals has the rules.
• GIFs in collab chats. A GIF sent to a room moves, at its own size, so a live chat can have some fun.

Improvements

• Every error code explains itself. The code at the end of an error is now a chip: hover it, or tap it on a phone, and it says why it happened and what to do, with a link to its row on the error codes page.
• One sound at a time. Play a song, a voice or a clip and whatever else was playing pauses, in chats, in collabs and in your library. Read aloud does the same.

Bug fixes

• Adding credits to a room, making one or joining one now shows the code when it fails, like every other error.
• A collab run or reading short of credits now carries a code (W301, W825).
• A reading that is too long (W821) and a start frame that is too large (W605) showed their code twice; they show it once.
• Read aloud on a second reply while the first was reading stopped both; the second now takes over.

Also

• Docs. Every code on the error codes page has its own link, and W914 (the card checkout box) is listed. Referrals covers public collabs.

1.45, Voice, music, sound, 2026-09-25

Audio is the fifth door. Have a voice read your words, ask for a song, or describe a sound and get it, in chats and in collabs, priced per character, per song or per second like everything else.

New features

• Voice. Twelve voices read what you type: ElevenLabs Turbo 2.5 by fal at 50 credits per 1,000 characters, ElevenLabs v3 by fal at 100, and ElevenLabs Flash 2.5 and v3 through the gateway. Up to 5,000 characters a reading (3,000 on v3).
• Music. Describe a song and get one with lyrics: Lyria 3.5 by fal at 100 credits a song, Lyria 3 Pro at 80, and ElevenLabs Music by fal at 600 credits a minute for a piece between 30 seconds and five minutes.
• Sound. A door creaking, rain on a tent, a bell: ElevenLabs Sound Effects by fal at 2 credits a second, one to 22 seconds.
• In collabs too. An Audio pill in the room composer; @askr with the words or the description and the room hears it, paid from the room balance, kept in the room library.

Improvements

• A player on the line, lyrics under a song, a Made with AI line under every piece, Download on the receipt, and the library keeps your audio beside your pictures and clips.
• Priced before you send. The button says what a reading, a song or a sound will cost, and a refused request costs nothing.

Bug fixes

Nothing this time.

Also

• Docs. Make audio covers the models, voices, lengths and prices; the codes W820 to W829 are on the error codes page.
• Privacy. The privacy notice says what fal receives for a reading, a song or a sound: the text or the description and the name of the voice, kept under fal's terms, used to train nothing.
• API. Audio is not on the public API yet.

1.44, Public collabs, 2026-09-25

A collab can now be public: share the link and anyone can watch it live, no account needed, and anyone with an account can join, up to a hundred people on one balance. Private stays the default, and the creator now decides what a room can do.

New features

• Public collabs. On the People card, or when you start a room, pick Anyone with the link. The room, its answers as they stream, its pictures and clips and its library are then open to anyone who has the link, signed in or not, with a Join button for those who are. Invite only puts it back the way it was; whoever joined stays.
• Join from the link. A public collab takes anyone with an account and an @, up to 100 people (a private one keeps 12). Past that the link is view-only. Sign in to join brings you straight back to the room.
• What a room can do. Room settings has a Can use row: the creator turns Ask, Code, Image and Video on or off for the room, and the composer offers only what is on.
• Room settings for everyone. Every member, and every viewer of a public room, sees the settings and the room balance; only the creator changes them or adds credits.

Improvements

• A shared link looks like the room. A collab link unfurls with the askr mark and the room's title on the card; a private room's card names no title. Every other page of heyaskr shares with the brand card too.
• The room's side. Room settings, then Room (the threads, renamed), then People, from the top of the screen and building downwards as settings open, threads are made and people come in.
• Copy link beside the choice, and the New room form asks who can see it.

Bug fixes

• Public, and how many are watching, on the room's title line; the phone's sheet offers Sign in to a reader with no account.

Also

• Privacy. The privacy notice says what public means for a collab: anyone with the link can read it, watch it live and play what was made in it, without an account.

1.43, Direct Anthropic routing, 2026-09-25

Claude now runs on Anthropic's own API. Pick the version marked (direct) under any Claude model and it thinks as long as it needs, shows its reasoning as it goes, costs Anthropic's list price, and keeps the conversation cached between turns so a follow-up costs a fraction to read. The gateway versions stay where they were.

New features

• Claude (direct). Opus 5, Sonnet 5, Fable 5.1, Opus 5.5 and Haiku 4.5 straight from Anthropic, listed under each Claude model as its (direct) version, in chats and in rooms.
• Room to think. A direct reply can run to 64,000 tokens, thinking included, and the reasoning streams while it happens instead of a silent wait. The handoff to another model when Claude ran out of room no longer applies to these versions.
• Cached conversations. The transcript is cached between turns, so a follow-up reads the history at a tenth of the input price or less. On a long chat that takes most of the input cost off.
• List price. A direct version costs what Anthropic charges, 5.5% under the gateway's copy on every token.

Improvements

• Receipts name the source. A run on a direct version says (direct) on its receipt, and the ledger records where it ran.
• Native PDFs. A direct version reads a PDF of up to a hundred pages as the file itself, and pictures as pictures.
• If Anthropic is busy or unreachable, the turn runs on the gateway's version at its price, and the receipt names the one that ran. A refusal never does.

Bug fixes

Nothing this time.

Also

• Content rules. An account that a direct version refuses five times in a day loses the direct versions for a day (W416, on the error codes page). Every other model keeps running.
• Privacy. The privacy notice and the docs privacy page name Anthropic behind the versions marked (direct): it receives the request with a fixed code made from your account id, keeps it for 30 days under its commercial terms and does not train on it.
• API. The direct versions are not on the public API; its Claude models run through the gateway as before.

1.42, Video by fal, 2026-09-25

Seven versions of the video models you already use, served by fal at its list price: Kling 3.0 at less than half today's price, Veo 3.1 in place of Veo 3, and Wan 3.0 and Seedance 2.0 at about 13% less. Pick them in the video picker under the model they belong to, labelled "by fal". Same models, lower price, nothing else changes.

New features

• Kling 3.0 at less than half price. Kling 3.0 Pro by fal renders a 5-second 1080p clip for 840 credits instead of 1,932, and Kling 3.0 Standard for 630 instead of 1,449. The 10 and 15 second clips are priced the same way.
• Veo 3.1 is here. Veo 3.1 and Veo 3.1 Fast come in as versions of Veo 3: an 8-second clip with sound for 3,200 credits instead of 3,680, or 1,200 instead of 1,380 on Fast.
• Wan 3.0 and Seedance 2.0 by fal. About 13% less than today: a 5-second Wan 3.0 clip for 500 credits instead of 575, Seedance 2.0 for 1,520 instead of 1,740 and Seedance 2.0 Fast for 1,210 instead of 1,390.
• Sound on every clip, and a start frame on all but Veo 3.1 Fast, with the same lengths and shapes each model offers today.

Improvements

• The price you see is the price you pay. A fal render is never charged above the quote the picker showed. If fal bills more, askr absorbs it.
• A queued render keeps its hold. fal queues work when it is busy, so a render keeps its credits held for up to two hours while it waits, instead of the hold lapsing at twenty minutes.
• If fal cannot take the job, the render runs on today's route at today's price, and the receipt names the model that ran.
• Clearer names in the picker. A supplier's newer model sits under the model it replaces by its own name, as Veo 3.1 by fal under Veo 3, and a version that only starts at a longer length says what the model charges for the same clip.

Bug fixes

Nothing this time.

Also

• Privacy. The privacy notice and the docs privacy page name fal behind the versions labelled with its name: it receives the request, with your prompt and any start frame, and keeps it for 30 days and the clip for at least 7. It makes no zero-retention promise, so pick the plain version when that matters.

1.41, Pay by card, 2026-09-24

Add credits by card, Apple Pay or Google Pay as well as crypto. The price you pick is the price you pay, tax included, and the credits land as soon as the payment clears. Card credits are for your own use: to keep fraud and money laundering out, they cannot be sent, cashed out or turned into crypto.

New features

• Add credits by card. Top up opens on Crypto with Card one tap away, and Balance & deposits now has Add credits by crypto and Add credits by card. Pick $5 to $100 or your own amount, tick that you want the credits now, and Checkout opens Stripe's secure page. A dollar is 1,000 credits, the same as crypto, and the smallest payment is $5.
• The ways to pay where you are. Cards, Apple Pay, Google Pay and Link work everywhere, with local options such as Klarna, Pix and Samsung Pay depending on your location. The price shows in your own currency.
• Tax included. The price you pick is what you pay: any VAT or sales tax is inside it, and askr adds no fee.
• Credits land at once. Back on askr the balance fills as soon as Stripe confirms the payment, and a confirmation email follows. Card credits get the $ASKR holder discount, like any credits you buy.

Improvements

• Card credits are for your own use, on purpose. We are against fraud and money laundering, so credits bought by card follow the rules below. They keep a stolen card, or money that needs hiding, from passing through askr. Credits bought with crypto work as before.
• They cannot be sent to anyone. Send credits refuses them, so a stolen card cannot be turned into credits in someone else's account. Your balance shows how many of your credits came from card.
• They cannot be cashed out or turned into crypto. No credit on askr can be withdrawn, so askr is never a route from card money to cash or coins.
• A collab room is not a way round it. Card credits can fund a room you share, and whatever the room hands back returns to you as card credits, still unsendable.
• No chargebacks on credit you used. Used credit is not refunded to the card, and a chargeback on it is contested with the account's usage record: the purchase, your agreement at checkout and the requests that spent it. A run that fails is refunded in credits, as always.
• A dispute pauses the account. While a card payment is disputed, the account cannot send credits or buy more by card. If the dispute is lost, credits from that payment come back off the balance, and anything already spent is owed before the account can run models again.
• No referral bonus on card purchases. Signing people up with stolen cards earns nothing.
• askr on Stripe's page. The checkout carries askr's name, logo and colours, and your card statement reads LINK.COM* HEYASKR.AI.

Bug fixes

Nothing this time.

Also

• Terms and privacy. Card purchases are sold through Link, Stripe's merchant of record. The terms now cover card purchases, refunds and disputes, and the privacy notice names Stripe and Link. Add credits in the docs covers paying by card.

1.4, askr Skill, 2026-09-24

Ten skills behind the Ask tile, each a chat with its own screen that builds itself around what you ask, and starts before the answer lands. The model writes plain words and small pieces of structured data; your device draws them as cards, calculators, steps and sources, checks them, and does the maths.

New features

• Ask opens a chooser. Tap Ask and pick Ask anything, the plain chat as before, or one of ten skills: Look things up, Write and edit, Learn and homework, Fix it and how-to, Health and fitness, Translate, Money, Brainstorm, Summarise, Shop and compare. Each opens a chat built for that job, on the model that suits it, with its own message box hint and its own settings.
• A screen per skill. Beside the chat on a desktop, a sheet on a phone, each skill draws what it made. Look things up shows every page it read as a card with the exact words that back each sentence. Write and edit keeps the draft beside the chat, with the changes marked, a word count and reading age worked out on your device, and one-tap rewrites. Learn and homework reveals a worked solution one step at a time with a hint on each, checks your arithmetic on the device, and marks a quiz there too. Fix it shows a Fix card: what you need before you start, the difficulty, the steps easiest first, a did-this-fix-it check under the current step, and any risky step locked until you have read its warning. Health and fitness draws each test result on a bar against the range printed on your report, builds your appointment list to tick off on the day, and shows a care card that says when to get help. Translate gives a two-part card with tappable words and other choices, a clarifying question when the meaning could go two ways, and photo translation for menus and signs. Money runs loan, mortgage, savings, debt and budget calculators on your device with the maths shown, and reads a statement photo into spending by category and the charges that repeat. Brainstorm deals ideas as cards to keep, drop, combine and build on, with a lens to filter them. Summarise gives an overview card with the quote behind each point at the length you choose. Shop and compare asks what matters first, then puts the options side by side with the better one marked on every row.
• It starts before the answer. The moment you send, the panel opens with the skill's real shape, empty and honest: the headings, the rows, the controls, and one line on what is coming. Nothing is invented; the only content is your own words. Ask about a mortgage of 250k at 4.5 percent over 25 years with a 30k deposit and the calculator opens with those numbers already in it, free, while the model is still thinking. Say what is broken and the Fix card appears with its steps to come. Ask a question and the pages show up as askr reads them. The ghosts shimmer while a reply streams and hold still when it ends, and on a phone a slim bar appears at send, pulses while askr works, and opens the sheet.
• Price, Balanced or Performance. Sits in the message box where the mode chip was, beside the model it runs. Price runs a cheaper model with short answers, Balanced is the everyday middle, Performance runs the strongest models, asks for thorough checked answers, and in Health, Money and Fix it also searches the web. Tapping a tier switches the model; picking a model by hand moves the tier to match, or to Custom. The choice is remembered for that skill's next chat. There are no cost estimates inside a skill: the receipt under each reply says what the turn cost.
• Your settings as pills. Each skill's own settings sit over the message box as one kind of pill: your device and its system, the languages and the tone, the level and whether askr guides or explains, your country and currency, what you need from a purchase, the length and format of a summary. Menus and panels float over the chat instead of pushing it down, and the settings carry to your next chat in that skill.
• Free on your device. The calculators open at 0 cr with example numbers. The arithmetic check in Learn, the quiz marking, the word diff in Write, the lens in Brainstorm and the reading level all run on your device, with no model call and nothing charged.
• Double-check with another model. In Look things up, Learn, Health, Money and Summarise, one tap asks a model from a different family to look for mistakes in the last answer and say what it would change. Only when you tap it, priced as one turn on that model.
• Photo first. Fix it, Health, Translate, Money and Learn put a camera button first: a photo of the label, the report, the menu, the statement or the question. Summarise and Health also take a PDF or a document.
• Save or share. Whatever the skill made, the draft, the steps, the appointment list, the shortlist, the comparison, the sources, copies as text, saves as a file, or goes to your phone's share sheet.
• A skill offered from plain chat. When a plain chat asks about a loan, a symptom, a translation or a fix, a chip offers to ask it in that skill with the question ready to send. Decided on your device from your words, no model call, easy to dismiss.

Improvements

• How it works. Each skill's instructions are held on our server, added to every turn of that chat with the date in your time zone, your language and your settings, and no client can write them. The model answers in plain words plus fenced blocks of structured data with a kind, such as steps, draft, labs, calculator or citations. Your device cuts the blocks out of the reply, draws each with the skill's own component inside its own error boundary, so a bad block costs one card and never the chat, and shows the rest as ordinary markdown, with maths typeset.
• What is checked, and where. A citation in Look things up must name a page the search actually read for that answer; one that does not is repaired to the real page behind the same number, without its quote, or dropped. A quote in Summarise must be words from the text you pasted. The maths in Money, the arithmetic in Learn and the diff in Write run on your device, never in the model. If your words describe an emergency, Health shows the care card on your device before the reply and tells the model to agree with it. Nothing is trusted just because a model said it.
• Live reference data. Health reads the matching MedlinePlus topic and Money the Bank of England Bank Rate, UK inflation, exchange rates and, when you ask, crypto prices, fetched before the answer, cached, with a short deadline, so a slow source never holds a turn up. Fetched text is wrapped as data the model is told to quote, not obey.
• Every model, one contract. Because the blocks are plain fenced JSON, every model on askr can run every skill; each skill simply opens on the model that suits it, a reasoning model for Learn and Money, a fast one for Translate and Summarise, the strongest for Performance.
• One size everywhere. The skill header is one slim row in every skill, the settings are 30 pixel pills, every label is in the site's own face and sentence case, and everything a skill adds over the message box sits on the message box's own column at any screen width.
• Search turns always answer. Before the last call of a search turn, the one with no tools, the model is told in plain words that no more searches can run and to answer from what it has. Without it, one turn in fifteen on a fast model asked for another search and wrote nothing. No extra call, no extra cost, and plain chats get it too.
• The message box. The Compare models side by side button is gone from the message box; the Compare page itself remains. The cost of a turn shows once, beside send, or in the model pill on a phone.

Bug fixes

• A shared chat link opened on the home. A link with a chat in its address, opened in a fresh browser, landed on the home page instead of the chat. The address was read after the first network wait, by which time it had already been cleared.
• Reopened search chats. A chat that searches now shows the globe lit when reopened, as it did when it was started.

Also

• Prompt tests. All ten skills were run at all three tiers against a judge that sees the history, the settings and the sources, and every block those runs produced was drawn through the real screens on a server. The test scripts ship with the code.
• Testing on staging. Hannah built the ten skills on an isolated branch and tested them on staging; the polish, the frames and the release went through staging before production the same day.

Announced on X.

1.395, Time to think, 2026-09-24

The Claude 5 family thinks before it writes, and on askr that thinking used to arrive as silence: the chat gave up after three minutes, then the connection upstream gave up at five. Claude now thinks where askr can see it, the chat says what is happening while you wait, and a brief Claude cannot finish is handed to a model that can.

New features

• Claude thinks out loud. Claude Opus 5.5, Sonnet 5, Fable 5.1 and Haiku 4.5 used to think in silence: not a byte arrived until the answer began, and after about five minutes of that the connection was closed upstream (code W511). Every Claude call now takes the route that streams the thinking as it happens, so the chat shows thinking from the first seconds and a long think is no longer cut short.
• Said before it happens. With a Claude model picked, the composer says it thinks before it writes and a big brief can take several minutes. Once a Claude reply has been thinking for 45 seconds, the card says so in full: that it may run out of room, that nothing is charged if it does, and where the brief goes next.
• Handed on, not dropped. A reply, thinking included, may be up to 32,768 tokens. When Claude thinks through that whole budget and never writes (code W514), nothing is charged and the same brief runs again on GPT-6 Sol as its own answer, under its own name, without you asking. A direct line to Anthropic is being built into our router so Claude gets the room it needs.

Improvements

• Ten minutes to think. A model has ten minutes before its first word and three minutes between words once it is writing; an answer may run for twenty minutes in all.
• With Search the web too. The wait after the searches, where the model reads the results and thinks, has the same limits and the same notes.

Bug fixes

• Stopped after one sentence. With Search the web on, three minutes of quiet was taken for a dead connection and the answer stopped with whatever had arrived. Those turns were charged only for the sentence that arrived, and where an allowance covered them, nothing came off the balance.

1.39, File attachments, 2026-09-24

Give any model a file. Pictures, PDFs, text and code, Word, PowerPoint and Excel, audio and video go in from the paperclip, in the app, in Collab rooms and on the API. A model that cannot read a file itself gets its words, so a PDF works on every model.

New features

• Attach almost anything. The paperclip, a drop or a paste takes pictures (PNG, JPEG, WebP, GIF, up to 8 MB), PDFs (32 MB), text and code files (5 MB), Word, PowerPoint and Excel documents (32 MB), audio (MP3, WAV, M4A, OGG, FLAC, 25 MB or 20 minutes) and video (MP4, MOV, WebM, 100 MB or 10 minutes). Up to 6 pictures, 3 PDFs, 5 text files, 3 documents, 2 audio files and 1 video on a message. Uploading is free; a file costs tokens only on the messages that carry it, and the estimate and the receipt both say so.
• PDFs and documents work on every model. A model that reads files gets the file. Every other model gets its words, read once on the way in, so the cheapest model in the picker can read your contract too. A PDF over 100 pages goes as text everywhere.
• Audio and video, on the models that take them. 32 chat models hear sound and 66 watch video, as the catalogue stood on 24 September. Audio for any other model goes as a transcript, made once and kept with the file.
• A Files page. Everything you have uploaded, for chats and Collab rooms, in one list: what it is, what was read from it, and when. Open, download or delete. A chat's files go when the chat goes; an upload never sent is cleared after seven days.
• In Collab rooms. Attach the same kinds in a room. The file is the room's, readable by its members, and stays with the room.

Improvements

• Pictures at full size for edits and video. A picture for an image model to change, or the start frame of a clip, now goes at up to 4,096 pixels on its longest side and 6 MB, not shrunk for chat. 212 chat models see pictures.
• Ask again about a file. Your three newest messages that carry files send them in full; older ones are named in a line, so a PDF from the start of a long chat is not paid for on every reply. Files count toward the model's window, and a turn that does not fit says so before it runs.

Bug fixes

Nothing this time.

Also

• For developers: POST /v1/files uploads once, then {"type": "file", "file": {"file_id": ...}} in any message, in OpenAI's shape, so client.files.create works as it is. image_url, file, input_audio and video_url parts go inline too, and inputs on GET /v1/models says what each model takes. Files.
• In the docs. Attach files has the limits, the per-model rules and the costs; codes W604 to W612 on Error codes say what a refused file means.

1.388, Credits and clips, 2026-09-24

Airdropped credits are spent first, then credits given to you, then the credits you bought, and every receipt says which. Runway Gen-4 no longer offers a 10-second clip the gateway does not deliver, a clip shorter than the one you paid for is charged at the shorter price, and the render confirmation is askr's own.

New features

Nothing this time.

Improvements

• Spend order, spelled out. Free (airdropped) credits go first, then credits given to you (a referral bonus, a transfer in, an operator's credit), then the credits you bought. Every receipt and the balance panel say which pot a run came from.
• Charged for the clip that arrived. Every finished clip is measured before it is charged. Shorter than the length you paid for, on a model that prices the shorter length, and you pay the shorter price. askr eats the difference.

Bug fixes

• Runway Gen-4 renders 5 seconds only, for now. The gateway bills the 10-second price and returns a 5-second clip: on 24 September three request shapes all came back at 5.04 seconds. The 10-second option is withdrawn until it works, and everyone charged for it is refunded.
• Render it? is askr's own dialog. The browser's confirm box could be silenced by the browser after a few renders in a row, and the Render button then looked dead until a reload. It cannot happen with ours. The same for the prompt when you edit a message that has replies.

1.387, Scroll while writing, 2026-09-24

Scrolling up during a fast answer no longer pulls you back to the bottom. The chat follows an answer until you take hold of it, and follows again when you scroll back down.

New features

Nothing this time.

Improvements

Nothing this time.

Bug fixes

• Take hold and it stays. Since 1.383 the chat followed an answer only while you were near the bottom, but a fast model wrote quicker than you could scroll out of that zone, so a flick was undone within a token or two. The chat now stops following the moment you scroll up, however fast the answer arrives, and the answer keeps growing below you until you come back for it.
• On a phone too. A flick that left the transcript still moving when your finger lifted used to count as being at the bottom, and the next token dragged you back. The direction you moved decides now, not where you happened to be when you let go.
• Back down, it follows again. Scroll to the bottom and the chat follows the answer as before. A new message of yours still goes to the bottom, and so does opening another chat.

1.386, Free credits first, 2026-09-24

Airdropped credits are spent before your own balance, on every model. The balance panel in chat now shows exactly that: an answer paid from free credits moves the free line, not your balance.

New features

Nothing this time.

Improvements

• Spent this session counts both. The line under the balance adds what the free credits covered to what the balance paid, so it is what the session cost, not only the part you paid for.

Bug fixes

• Free credits go first, on every model. A turn, a picture or a clip comes out of your airdropped credits until they are gone, and only then out of your balance. That has been the rule since the first airdrop; the panel in chat was the part that got it wrong.
• The panel moves the right number. It used to take every charge off your balance on screen and leave the free line alone until you reloaded, so a session on free credits looked like it was spending your own money. Every receipt now says how much the free credits covered, and the panel takes that off the free line and only the rest off the balance.
• Read back after anything irregular. A stop, a reload mid-answer or a render checked later re-reads both numbers from the account instead of guessing.

1.384, The whole window, 2026-09-24

Every chat on askr, in the workspace, in collabs and on the API, now carries up to 25 times more per turn: up to about 3 million characters, sized to each model's own context window, which is about 1 million tokens on the largest models.

New features

• 25 times more per turn. askr used to send any model at most 120,000 characters of a conversation, about 40,000 tokens, however much the model could hold. Each turn is now sized to the model's own context window, with room kept for the reply, up to about 3 million characters. 351 of askr's chat models get bigger turns.
• Up to 1 million tokens. The largest models, Kimi K3, GPT-6 Sol, Claude Sonnet 5 and DeepSeek V4 among them, can now use their whole window of about 1 million tokens. A model with a 200,000-token window goes from about 40,000 tokens a turn to about 167,000.
• Whole documents, one question. One message can now fill the whole turn, where it used to stop at 48,000 characters. Paste a codebase, a book or a stack of reports and ask about all of it at once.

Improvements

• Long chats keep their start. Up to the last 400 messages are sent to the model, up from 20, so a long conversation no longer forgets how it began.
• In collabs too. A room now reads up to 400 messages of its thread and as much as its model's window holds, up from 20 messages and 100,000 characters, and one message can run past the old 24,000 characters.
• On the API too. /v1/chat/completions takes the same: up to 400 messages and the model's own window, up from 40 messages and 120,000 characters.
• Kept whole. A long chat is saved in full to your account and stays in your history, without pushing your other chats out of the browser.
• You pay for what you send. Tokens are priced as before, and a bigger turn holds more credits before it runs. A model with a small window keeps the limits it had.

Bug fixes

Nothing this time.

Also

• Tested at full size. A 779,000-character document, about 161,000 tokens and six times the old limit, went to Kimi K3 in one message. The answer came back in 20 seconds, with a detail from the very first line.
• In the docs. Workspace limits, API limits and code W106 now say so.

1.383, Room to think, 2026-09-23

Long code answers and big build briefs no longer run out of room. The reply ceiling is four times larger, a model that spends it all thinking now says so, a long answer is given the time it needs, and the chat no longer drags you down while it writes.

New features

Nothing this time.

Improvements

• Four times the room. A reply may now run to 32,768 tokens, thinking included, up from 8,192. Long code answers arrive whole instead of stopping mid-file. A thin balance gets the largest reply it can hold rather than a refusal, and a model that takes less than the ceiling is retried one size down on its own, so nobody sees a refusal for it. If an answer still reaches the ceiling, the receipt says so and "continue" carries on.
• Time for a long answer. The four-minute cap on an answer is gone. As long as the model keeps writing, the answer keeps coming, for up to twenty minutes, so a long file is no longer cut for time either.

Bug fixes

• A model that thinks itself out of room says so. Some models, the Claude 5 family among them, think before they write. When it used up the whole budget the reply said "did not answer, try again", which sent people round in circles. It now says what happened, with code W514, and suggests a model that thinks less. Credits go back either way.
• Read while it writes. The chat follows an answer only while you are at the bottom. Scroll up to re-read and it stays where you put it; scroll back down and it follows again. A new message of yours still goes to the bottom.

Also

• In the docs. The reply length and code W514 now say 32,768.

1.382, Send credits, 2026-09-23

You can send credits to another askr account. Bought credits and credits you were given move; airdropped promo credits and the holder discount stay where they are.

New features

• Send credits from your balance. In Settings, Balance & deposits, a Send credits button beside Top up: enter an @, an amount, a note, confirm, and both of you get a receipt and a line in your Sends list.
• What moves and what does not. Credits you bought or were given move. Airdropped promo credits keep their expiry and never move, and the holder discount follows your wallet.
• Limits that keep it safe. Minimum 100 credits, credits deposited in the last 24 hours wait a day, and a six-digit email code for larger sends.
• Final, with a safety net. A send is final once confirmed. If one was made by mistake or by someone who was not you, tell us and we can reverse it while the credits are still there.

Improvements

Nothing this time.

Bug fixes

Nothing this time.

Also

• Still spend only. Sent credits buy model usage like any other; nothing here turns credits back into cryptocurrency or cash. The terms say so. See Send credits.

1.381, Search the web, 2026-09-23

Any model can now search the web and read pages when it needs live facts. Off by default; a globe in the composer turns it on, and every search is priced before you send.

New features

• A globe beside attach. The model may search when your question needs something current. The globe glows while a search runs and shows how many ran. Remembered per chat, with a setting to have it on for every new chat.
• Priced before you send. The estimate adds the most a turn can spend on searching, 5 credits a search plus the page text at the model's rate. The receipt shows what actually ran, for example 2 SEARCHES, 10 CR. At most 3 searches and 2 pages a turn.
• Sources on every answer. Numbered [n] marks in the reply link to a Sources line underneath, so you can check what the model read.
• It knows when to look. Ask about today, a price or the latest news with the globe off and the search switches itself on for that chat, the globe glows, and the receipt shows what it cost. Turn the globe off in a chat and it stays off there.
• In Collab rooms too. A globe in the room composer, per room. An @askr line that needs live facts searches, the globe glows for everyone in the room, the reply carries its sources, and the searches come off the room balance like the rest of a room's model use.

Improvements

• Only where it works. The globe is greyed out on models that cannot use tools, with a note saying so. Zero-retention versions of a model search too.

Bug fixes

Nothing this time.

Also

• For developers: `web_search: true` on chat completions, streaming or not, with sources returned as url_citation annotations and the searches in usage; supportsTools on GET /v1/models.
• Privacy: with Search the web on, the search terms the model chooses go to Brave Search; pages are read by askr's server, never from your browser. See Search the web.

1.38, New chat UI, 2026-09-23

The chat, the collab rooms and everything around them get a new look: the moon and its falling light, a rail that opens into a sidebar, glass surfaces, Geist type, and a light mode beside the dark one. Every result is still priced before you send it.

New features

• A new look for the whole app. The chat at /chat, every collab room and the pages around them are redrawn: the red moon with its falling light lines on the home, a 64px icon rail that opens into a 264px sidebar, glass chips for the $ASKR price, your balance and your account, and Geist type with every number, price and label in Geist Mono. On a phone: a top bar, a menu drawer, and a composer docked above the keyboard.
• Light mode. Choose Light, Dark or System from the sun/moon switch in the rail, the sidebar, the phone menu and the Collab room, from your account menu, or in Settings under Appearance. The app opens in Dark unless you choose otherwise, your choice never flashes the wrong theme, and the site and the docs stay dark.
• A home that gets out of the way. The askr logo, then one line, "What are we building?", whose last word rolls through more than seventy words. Five mode tiles in one row: Ask, Code, Images, Video and Collab. The composer is glass, grows as you type, shows the model and its price before you send, and takes attachments, compare and dictation. Recent chats and your Library sit below it. The Audio and Skills tiles, the old welcome line and the footer line are gone.
• The model picker. Cmd-K opens a dialog on desktop and a bottom sheet on a phone: search, the pills All, Ask, Code, Image and Video, A to Z and price toggles, keyboard from top to bottom. All lists every model on the platform, Ask the popular and cheap general ones, and Code every text model that can write code, with the code-flavoured ones first.
• Top up without leaving the chat. Top up opens as a dialog or a sheet from the balance chip, the sidebar card and the phone balance pill: pick the asset from the ones the processor offers, the amount, and you get the address and a QR that scans in both themes. A deposit you start and close is watched to its end and lands on your balance.
• History, Search, Library, Collab and Settings as pages of their own. History groups your chats by day with mode filters, what each chat cost, Rename and Delete. Search covers your chats and your library with highlights. Library shows your pictures and clips as tiles with Download, Re-run, Open chat and Delete. Settings holds your account: General with Appearance and a default model per mode, Balance and deposits, $ASKR benefits, API keys and Privacy. /account now lives at /chat/settings.
• Edit forks the chat. Edit a prompt you sent and it goes out again from that point as a new, charged turn, while the old words and the replies they got stay as an earlier version. A control beside the turn flips between versions.
• Collab rooms. The room title is a dropdown with a Switch room popover, the chat column sits beside cards for People, Threads and Room settings, and a phone gets sheets for each. Every member sees the room settings and the room name; only the creator changes them. Reactions sit on the bubble they belong to. A room that cannot load says so and offers Try again. "Collab rooms" is now just "Collab".

Improvements

• Calm inside a chat. Once a conversation starts, in any mode, in a collab room and on the pages, the light lines stop and the moon dims, so nothing moves behind what you are reading. The home keeps its motion. Reduced-motion settings are respected everywhere.
• Replies with receipts. Every answer carries the askr mark, the model, the mode and its status, and a receipt line that starts with the price: Copy, Read aloud, Re-run, Compare and Download as plain links. Code blocks have a header with the file name, diff counts and tinted added and removed lines. Pictures show two up, compare shows four cells with Pick, clips show on a card with a rendering row while they render, and Read aloud plays in a small player that fits a phone.
• Chats run on their own. Each open chat keeps its address, like /chat?chat=abc, so a reload or a link comes back to it, and /chat on its own is the home. A reply still writing in one chat no longer blocks sending in another, and Stop only stops the chat on screen. Home always opens a fresh chat.
• Everything from 1.37, in the new look. Failure cards that say why and whether credits went back, pictures and clips shown whole at their own shape, one-click copy, editing a sent message and deleting from your library are all carried into the new chat and collab rooms, and a reader at the foot of a chat stays there when a picture loads.
• Built to be used with a keyboard and a screen reader. Every control is reachable by Tab with a visible red focus ring, dialogs trap focus and close on Escape, the phone drawer makes the page behind it inert, fields show focus on their rounded box, hit areas are at least 44px on phones, and text meets WCAG AA in both themes.

Bug fixes

Nothing this time.

1.375, Second supplier, 2026-09-23

Many models are now served by a second supplier, Engy, at a lower price and with zero data retention. Pick the version you want; nothing else about how askr works changes.

New features

• The same model, cheaper. In the model picker a model served by both suppliers shows a second row beneath it, "by Engy", with its own price. Pick either.
• Eight models to start. The saving is against the same model on askr before, measured on a standard turn. DeepSeek by Engy: V4.1 Flash 86% less, V4 Flash 84% less. GLM by Engy: 5.2 62% less, 5.3 34% less, 5.3 Flash 14% less. Kimi by Engy: K3 38% less. Qwen by Engy: 3.8 27B 90% less, 3.6 35B 71% less.
• Zero retention on those rows. Every Engy version except one is served on hardware Engy operates, with prompts and answers never stored or trained on. The row carries a ZERO RETENTION chip and the model button says so.

Improvements

• Priced the same way as everything else. The estimate beside Send, the receipt after a reply and the holder discount all work as before; the receipt says BY ENGY when that version ran.

Bug fixes

Nothing this time.

Also

• For developers: `GET /v1/models` lists the new versions as engy/<model> ids, and /v1/chat/completions accepts them, streaming included.

1.37, Feedback fixes, 2026-09-23

Fixes and improvements from what people told us. When a model says no, askr now tells you why and that you were not charged, instead of a failure that looks broken.

New features

• One-click copy on everything. Every prompt and every answer has a Copy, in the workspace and in a collab room: words copy as text, a picture copies as the picture itself where your browser allows, and a clip copies its link.
• Edit what you sent. In the workspace, edit a prompt and send it again from that point, after a check that says how many messages it replaces; what those messages made stays in your library. In a collab room, fix the words of your own message, and everyone sees it marked edited.
• Delete from your library. Every picture and clip in your library now has a Delete that asks first, so nothing goes with one stray click. A deleted item is gone at once; only the billing record of what it cost stays.
• Live prices on the homepage. The card on the homepage now shows the newest Claude, the newest ChatGPT and the most used model from the next most popular brand, each with its price and the $ASKR holder price, which applies within your daily allowance.

Improvements

• Pictures and videos shown whole. In a chat or a collab room, a generated picture or clip now shows its whole frame at its own shape, sized to fit, instead of a square crop until you click it. Reading further up while one loads no longer pulls you to the bottom.
• The same message every time. You see the same reason whether you wait on the page or come back later, and never the provider's raw text.

Bug fixes

• When a model says no, you hear why. A request a provider refuses under its content rules, like a video of a real public figure, now says so in plain words, tells you nothing was charged, and suggests rewording it or picking another model. Codes W816 for video, W415 for chat and images.
• No more "try again" for something that cannot work. A refusal never tells you to retry unchanged, because the same prompt would be refused again. Rules differ by provider, so another model may accept it.
• Videos no longer hang on a refusal. A render the provider blocks is marked failed as soon as the provider says so, instead of sitting on "taking longer than expected", and a brief network blip while we check on a render no longer fails one that is still going.
• A failed job looks like a failure. In the workspace, a request that fails now shows a failure card with the reason, and says whether your credits went back or nothing was charged, instead of reading like an answer.
• Long videos keep their place. A clip still rendering after 7.5 minutes, or through a dropped connection, keeps being watched and lands on its card when it finishes, even after a reload.
• In a collab, a clip that fails says so, and its reserved credits go back to the room, and a refusal never tells the room to try again.
• Footer: the chain badge, email and copyright now line up on one edge, on desktop and on phones.

Also

• Every code explained: workspace error codes now lists W415 and W815 to W819.
• For developers: the API answers a refusal with 400 and the code content_policy_violation, before or during a stream. See API errors.

1.36, ASKR Connect, 2026-09-22

Link the wallet that holds your ASKR without connecting it, signing anything or sending a single token. Paste the address, buy a little ASKR from that wallet in the app you already use, and the chain shows it is yours. Built for FOMO and every app that can buy but cannot connect.

New features

• Two ways in. Connect wallet signs a message, as before. Verify holdings is new, and needs nothing but your address.
• No connection, no keys, nothing sent. Paste your address, press start, and buy any amount up to 25,000 ASKR from that wallet within sixty seconds. Tap Done any time in the next ten minutes and the wallet is linked.
• Works from FOMO. A FOMO buy is made by your own FOMO wallet, so it proves the wallet is yours. Not sure of your address? The page shows you where it is: in FOMO, Account, then Deposit, then Deposit crypto.
• Only you can do it. Anyone can send tokens to an address, so a balance going up proves nothing; a buy the wallet made itself does. Tokens someone else sends you never count, and neither does a buy made before you pressed start.
• The same benefits either way. A wallet linked by a buy gets the same discount and the same access as a signed one, with its first allowance at the next 00:00 UTC.

Improvements

• Holdings stay fresh. A linked wallet's balance is re-read within ten minutes of any ASKR moving in or out of it. The daily allowance is still set once, at 00:00 UTC.

Bug fixes

• pump.fun and other apps that buy through a third-party router. A buy that reaches your wallet through a router cannot show the wallet is yours, so the page now asks you to sell the ASKR you just bought, and that sale confirms it.

Also

• Both ways, step by step: Holding ASKR.

1.35, Holder discount, 2026-09-21

Hold ASKR, link the wallet, and every model on askr costs you a twentieth of its price on the credits you bought. How much you get at that price each day scales with how much you hold: 1,000 credits of AI a day for every 0.01% of supply, capped at 25,000. It resets at 00:00 UTC.

New features

• Every model at 5% of its price, not a fee discount: askr is feeless, so the 95% comes off what the model itself costs and askr covers the rest.
• On the credits you bought. The holder price applies to credits you deposited, and the allowance counts only those. Airdropped credits and credits given to you are spent first, at the ordinary price, so a free credit is never doubled by the discount.
• Your allowance scales with your holding: 1,000 credits of AI a day for every 0.01% of supply you hold, capped at 25,000 a day. Past the allowance you pay the ordinary price, so nothing ever costs you more than it would without the token.
• Read once a day, at 00:00 UTC. At that instant a bag of tokens sits in exactly one wallet, so it earns exactly one allowance wherever it moves afterwards. Splitting a holding across wallets earns the same as holding it whole.
• A wallet you link today earns its first allowance at the next reset, which is what stops anyone linking, spending, moving the tokens on and linking again.
• Unused allowance does not roll over. It is the size of the discounted lane for the day, not a grant.
• Shown everywhere money is: on your balance, in the workspace rail, on the wallet panel and inside a collab room, with the discounted credits left today and the time it resets. It shows at 0 balance too, so you can see what a deposit would buy.
• It works inside a collab too: a run in a room is charged against the runner's own allowance first, and the room's purse pays whatever is left. An allowance is per person per day either way, so a room of holders is simply cheaper to run.
• On the price list: every rate on /pricing now carries the holder price beside it.

Improvements

Nothing this time.

Bug fixes

Nothing this time.

Also

• You still need credits. The discount makes the AI cheap, not free: a balance is what the 5% is charged against, and a deposit starts at $5.
• The docs say the real rule: Holding ASKR replaces the old promise of 25% off a platform fee that was never charged.

Announced on X.

1.31, Collab fixes, 2026-09-21

What the first rooms needed: pictures and replies and reactions inside a room, code that reads the same in it as out of it, and the joining bug that stopped some invites working at all.

New features

• Talk as well as prompt: attach pictures, reply to a particular message, and react with an emoji. The picker opens where you pressed it, not in the corner.
• Code reads the same everywhere: coloured, labelled by language, one click to copy, inside a room exactly as outside it.

Improvements

• Your @ is where your picture is, on the profile, instead of buried in the referrals card.

Bug fixes

• Join actually joins. An invite link could land you on a room you were not in. Fixed the day it was found.
• The privacy page says what a room actually does with what is written in it, and what is forwarded to a model. It said three things that were not true of collab.
• Crediting an account with no email works again in the admin. Accounts have been able to exist without one since 1.1, and the action still assumed every account had one.

Also

• Page views, counted: analytics on every page, named in the privacy notice and the policy rather than added quietly.

1.3, askr Collab, 2026-09-20

An invite-only room where a few people work the same models together out of one purse. Chat, code, image and video, the same four doors as the workspace, with everyone in the room seeing every answer as it arrives.

New features

• A room with its own purse. Whoever opens it funds it, and every run in the room is paid by the room. Nobody spends their own balance by being invited.
• Invite-only, by link. The room is not listed and not searchable; the link is the door.
• All four doors inside it: chat, code, image and video, on the same models as the workspace.

Improvements

Nothing this time.

Bug fixes

Nothing this time.

Also

• A beta, and it says so. Rough edges are known and being worked; tell us what breaks from the box on the roadmap.

1.25, Voice and roadmap, 2026-09-20

Speak to askr on any device and it goes straight to the model you chose; every reply can read itself aloud. Both run in your browser, cost nothing and send no audio to askr. The roadmap now lives on the site, with a box to ask for what you want built.

New features

• Dictation: a mic beside the paperclip in the composer. Tap to listen, tap to stop; sending stops it. Words show as they are guessed and settle as you go, and a pause to think does not end the session.
• Read aloud: a link on every finished reply, in your browser's own voice. Code is skipped with a word, links are read as their text, long replies are read in pieces so it finishes. One reading at a time; tap to stop.
• See the mic hearing you: four bars beside the mic move with your voice while it listens.
• The roadmap, on the site: heyaskr.ai/roadmap shows every version shipped and the plan ahead in order, and has a box that sends what you want built straight to the team.

Improvements

• The composer fits what you type: the box grows and shrinks with the message, typed or dictated. On a phone the text runs full width with the buttons underneath.
• The site asks for the microphone on the first tap, and if you have said no it tells you where the switch is on your platform.

Bug fixes

• Docs put right: deposits are credited in full (askr absorbs the processor's cut), overpayments are credited, chats live on your account, and referrals have their own page.

Also

• In your browser, not on our servers: Chrome, Edge and Safari carry the speech engine, so the buttons appear there and stay hidden in Firefox. No audio is sent to askr, nothing is recorded, and neither feature spends credits. The privacy page says so.

Announced on X.

1.2, Handles and referrals, 2026-09-20

Every account gets an @ and a referral link. Anyone who joins through it pays you for life: 10% of their top-ups in credits the moment they land, and 10% of their fees once fees are on. No ceiling.

New features

• An @ on every profile: a username account's handle is its @; email and wallet accounts can claim one. heyaskr.ai/r/<your code> is your link. How referrals pay.
• 10% of every referred top-up, paid to you in credits the moment their deposit credits. Automatic, once per deposit, never out of their balance.
• 10% of their fees, when fees are on. Accrued per run and paid hourly as credits. Nothing accrues while v1 is feeless; switching the fee on is the only change left for that to start.
• A referral link survives the sign-up: the code follows every door (username, wallet, email) and holds for 30 days.
• Support where things go wrong: @askrsupportbot in the chat rail, under any refunded reply, on the top-up card, in the account tabs and at the foot of sign-in.

Improvements

Nothing this time.

Bug fixes

• Wallet fixes: a wallet that is an account's only way in can no longer be unlinked; BNB pays from a wallet with a proper QR; WalletConnect opens one modal, not two; Robinhood Chain is the active chain on phones.

Announced on X.

1.1, Made for crypto, 2026-09-19

Sign in with a username and password or with a wallet, no email needed. Link the wallet that holds ASKR and the account shows what you hold. Top up the crypto way.

New features

• Sign in without an email: a username and password, or a wallet signature. An email can be added later for recovery.
• Wallet sign-in on any wallet, through WalletConnect as well as browser wallets.
• Holdings on your account: ASKR held, share of supply, your band and the fee discount it earns, read from the chain on link, on demand and daily.
• Top up the crypto way: ten coins as tiles with a search under them, a checkout card with the address in a box and a Copy button, and a tracker that follows the deposit from sent to credited.
• A receipt by email when a deposit credits.
• Support on Telegram: @askrsupportbot takes a ticket and relays the answer.
• Chats follow the account across devices; the history is named by what each chat asks.
• Drop a picture anywhere on the page to attach it; tap a generated image to see it full screen.
• The code panel lists files by name, one open at a time, with Save all.
• The token page carries the contract address, the Team Finance lock, and the Buy and Chart buttons.

Improvements

• Models that stop answering hide themselves from the picker instead of failing your request.

Bug fixes

Nothing this time.

Announced on X.

1.0, askr is live, 2026-09-19

heyaskr.ai is open: every major model behind one sign-in, priced at upstream cost, paid in crypto with no card and no subscription. ASKR launched on Robinhood Chain the day before.

New features

• Four doors from one balance: chat, code, image and video.
• Feeless v1: a dollar in is a dollar of model; the processor's and the gateway's fees are absorbed.
• The developer API at /v1, OpenAI-compatible, with per-key daily caps and askr.credits_charged on every response.
• ASKR on Robinhood Chain, launched 18 September, an access token with bands by share of supply.
• Docs at home on heyaskr.ai/docs, generated from the code, reading the catalog live.

Improvements

Nothing this time.

Bug fixes

Nothing this time.

Announced on X.

Before 1.0

The site as it was built up to launch.

September 2026

• Docs moved home. These pages replace the old docs site; every old link redirects. The reference is generated from the code and reads the catalog live.
• Token launch set. ASKR launches Friday 18 September on Robinhood Chain. The token.
• Clips show their first frame while a render runs and after it finishes.
• The library keeps what you make. Images and clips are saved with prompt, model and cost, and served to you only.
• The workspace counts on the model you chose. Comparing models is an option you take, not something that happens to you.

August 2026

• Four doors in the workspace: chat, code, image and video, from one balance.
• Video in the workspace with the exact price before the render and a hold that is released if it fails.
• The developer API at the root of the domain, so a base URL of /v1 works with every OpenAI client unchanged.
• The CLI: install, log in, talk.
• Per-key daily caps and askr.credits_charged on every response.