We tested every model on askr against its label.
Here is what we found.

askr gives you a few hundred AI models under their makers' names. If you pick Grok 4.7, you should get Grok 4.7. On 26 September a user published a report titled "Is ASKR selling the AI model it says it is? No." They had tested three of our models over five days: grok-4.6, grok-4.7 and glm-5.3.

Their test is simple and anyone can run it. A model only knows the world up to the day its training data ends, so they asked each model about well-known events from late 2025 and early 2026, then asked the same models from official sources. The official copies knew the events. Ours said they had not happened.

The report found a real bug. It was not the one it looked like: the model behind the label was the right one, answering as if its training had only just ended. Section 2 shows why. We are grateful to the user. This is the test we should have been running ourselves, and from now on we do.

We read our own code first. askr passes the model name you pick to our supplier unchanged, with your messages. In solo chat and on the API we add no instructions (only Search the web adds a line with its rules), we never fall back to a different model, and an unknown model name is refused, not replaced. We also add no date. That turned out to matter.

Then we asked the same three models the same kind of questions, one at a time, twice: once as the user did, and once with a single line in front saying what today's date is.

A model trained on data that ends in 2024 cannot name the 2025 Nobel Peace Prize winner or a model released in February 2026, whatever you tell it. Grok 4, from July 2025, could not either. These models can. And the date line cannot invent knowledge: Claude Haiku 4.5, taken straight from Anthropic, whose knowledge ends in early 2025, got the same line and still could not name a single one of these events.

What made them look older:

• No date. A model does not know what day it is. Without a date it tends to assume it is still early in its training, and it treats anything later as not yet happened. This is our part: askr sent the models no date. Official apps usually do.
• Hidden instructions on one route. On the route our supplier uses for Grok 4.7, about 1,240 tokens of instructions sit in front of every message before it reaches the model. We measured it with askr's code out of the path: a bare "hi" sent straight to our supplier counts 1,243 prompt tokens. We did not put them there, and they make the model guarded: in one run it gave Gemini 3's release date and details of the October 2025 Louvre theft, then said no pope had been elected in May 2025.
• Grok denies rivals' launches. Grok models often say GPT-5 or Gemini 3.1 do not exist while naming later events correctly. The report's questions were mostly about rivals' launches.
• Models are poor judges of their own name. Claude Haiku 4.5, taken straight from Anthropic, calls itself Claude 3.5 Sonnet. A Grok calling itself something else proves nothing.
• No script on Grok 4.6. The report says it caught Grok 4.6's hidden instructions, including "a knowledge cutoff of 2024-10". When we measured our supplier's route for Grok 4.6, a bare "hi" counted 19 prompt tokens: nothing of that size sits in front of it. Asked to repeat instructions it does not have, a model will often write out the kind it was trained with.
• GLM's family blind spot is Z.ai's own. Our GLM 5.3 does not know GLM-5 or GLM-4.7, but neither does GLM 5.3 FlashX served by Z.ai itself, asked the same questions with the date. And GLM-4.6 came out on 30 September 2025, so a model that knows about 18 November 2025 is not GLM-4.6.
• Tools were our gap. Our API does not yet pass tool definitions to the model; the API docs say so, and tool calling is on the roadmap. A model that is never shown the weather tool cannot call it, so it types the call out or says it will check. That one is on us.
• Cheap does not mean fake. We pay our supplier its price for the named model on every request. The holder discount and the free credits come out of askr's pocket, not from swapping in cheaper models.
• Different answers on different days. Our supplier can send the same model to a different host from one request to the next. The receipts in section 3 will show which host answered.

Two more things in the report were askr's, not a model's. "The conversation is too long" is askr's own error message, word for word: until 24 September we capped a conversation at 120,000 characters, and since release 1.384 each model gets its own full context window, which is why a longer document went through later. "The model returned no content" is our refund path, shown when a model spends its whole budget thinking; in our own first pass 16% of calls ended that way at a small budget, and a larger one fixed nearly all of them.

Our catalog lists 383 chat models (the picker hides the ones that are not answering), so we tested all of them on 26 September, askr's code out of the path, no search, no tools. Four rounds:

1. The user's quiz, word for word, on every model.
2. Three open questions per model about events a few months before its maker's stated cutoff.
3. For the 25 models that still looked old, the same questions again, with and without today's date.
4. For every model, one question listing 23 well-known events from March 2023 to June 2026, with today's date given.

The newest event a model gets right is where its knowledge ends. We compare that with where its maker says it ends. To know what a genuine model looks like, we ran the same tests on five Claude models taken straight from Anthropic's own API: their knowledge ends 3 to 8 months before Anthropic's stated cutoff. So a model is "as labelled" when the gap is 9 months or less. We only suspect a model when the gap is a year or more with nothing closer anywhere in its answers, and every suspected model then went to three independent reviewers whose job was to prove it genuine.

No model on askr was found to be a different model from the one on its label. 55 could not be settled either way; they are marked "cannot tell yet" below and stay under the nightly check.

A note on the method's limits. Genuine models are cautious about the last months before their cutoff and often under-report what they know, so a short gap proves nothing and a long one is only a reason to look harder. Where a maker publishes no cutoff we use its release date minus six months. A hidden prompt can make a genuine model answer like an older one, which is why we test with the date and look at the route as well as the answers. We would rather say "cannot tell yet" than guess in either direction.

Until this week, almost everything on askr ran through one supplier. That let us launch with hundreds of models in a few weeks. It also meant that what reached the model depended on choices made two layers below us, which we could not see and did not check. The instructions in front of Grok 4.7 are one of those choices. That is the bug behind the bug.

The fix is a router of our own: askr connects directly to the companies that make the models, one by one, and the supplier keeps only the long tail, behind the nightly check. We did not start this because of the audit. We already have one direct connection live: Claude (direct), a straight line to Anthropic's API, has been running since 25 September, and every Claude answer on it comes from Anthropic and says so. Grok comes next, directly from xAI, then GLM from Z.ai, then OpenAI, Google and DeepSeek. Each direct connection shows "(direct)" after the model's name and charges the maker's list price.

The rule from now on is short. A model on askr is the model on the label, served by its maker or by a route we name on the receipt, and checked every night.

askr is not an inference market. We are building a consumer app, and it is already live: one place to talk to every major model, with Skills that turn a chat into a job done, Collab rooms where a group works with the models together on a shared balance, and Mind, a memory that follows you across all of it, coming soon.

The models are the ingredients. We do not make them and we do not pretend to. What we owe you is that the ingredients are what the label says, that you can see where they came from, and that when something looks wrong, you hear the whole story from us, whichever way it comes out. This page is that story. The nightly check keeps it true from now on.

Here is how to test whether any model is the model on its label, the way we test ours now. Then the accusations made about askr, the proper test for each and what it showed, and why the original tests fell short.

The ladder we used, short enough to paste into any chat. Add today's date in the first line.

Norway also topped the 2018 and 2022 medal tables, so count item 10 only alongside other 2026 answers. The newest item a model gets right is where its knowledge ends; set that against its maker's stated cutoff.

• They never gave the model the date, so real models called real events future ones.
• They leaned on self-identification, which genuine models get wrong too.
• They compared against official models running inside other apps, with those apps' instructions, instead of like for like.
• They read askr's own limits, the message cap, the dropped tools and the thinking budget, as proof about the model.

What they got right, we have taken on: the missing date, the hidden tokens, the tools and the empty answers are all in section 3.

Paste this into any model on askr, then into the same model at its maker. Put today's date in the first line. Without it, a model that does not know what day it is will often call recent events future ones, and that is the mistake this page is about.

Question 6 is a hint, not proof: genuine models often name an older version of themselves. A model that cannot answer 1 to 5 even with the date is worth reporting, together with the maker's stated cutoff for it.

If you find something we missed, tell us at heyaskr.ai/feedback. We credit the finder and refund anyone affected.

Everything on this page started with one person telling us something was wrong. That is how askr gets better, and we pay for it: every bug you find and every report you send earns free credits on your account. It is how we shipped 27 releases in the week since we went live on 19 September, from Collab and Skill to file attachments, web search, card payments and the direct line to Anthropic. The whole list is on heyaskr.ai/roadmap.