Skip to content
Research

DeepAI Code said it was GPT-6 Luna. Its own response said Crouton Alpha.

On November 7 and 8, every DeepAI Code free-model response we got named the model we picked in one field and a free stealth endpoint in another. Here are the raw responses, field by field.

In short

  • We make Freebuff, which competes with DeepAI Code, so we’re showing the raw data rather than asking you to take our word for it.
  • On November 5, DeepAI told its followers on X: "Fast models free, no limits." The post has about a million views.
  • On November 7 and 8, 2026, we sent 33 requests to DeepAI Code’s nine free models. Every response put the model we picked in its model field, and every one also carried an endpoint label saying it ran on stealth/crouton-alpha, a free anonymous model on OpenRouter.
  • On a plain request to write an is_prime function, the model’s reasoning summary described "a conflict between public deployment name and developer-provided model name".
  • All 33 responses reported an upstream inference cost of 0.
  • OpenRouter has since revealed Crouton Alpha as an early version of Meta’s Muse Ember. DeepAI changed its routing by November 10 and has said nothing to its users about it.

When you pick a model in an app, you’re trusting the app to send your code to that model. On the weekend of November 7, DeepAI Code’s free models were sending back two different answers to "which model is this?" inside the same response. We spent a while working out what we were looking at, and this post walks through it with the responses themselves.

What DeepAI promised

DeepAI launched the campaign on X on November 5: "The best coding AI costs up to $500 a month. We made it free." The post names Claude Fable 5.1, GPT-6 Astra and Claude Opus 5.5 for the daily allowance, and says: "Fast models free, no limits." About a million people have seen it. The free fast models are the unmetered ones in the picker, such as GPT-6 Luna, GLM 5.3 Flash and DeepSeek V4.1 Flash. Two days later we tested them.

One response with two names in it

We picked GPT-6 Luna, one of DeepAI Code’s free models, and asked it to write an is_prime function. We sent no system prompt and asked nothing about its identity. This came back, trimmed for length:

{
  "id": "gen-1794094546-kpGAJVfB8DnDd5fseUAR",
  "model": "openai/gpt-6-luna",
  "provider": "DeepAI",
  "choices": [{
    "message": {
      "reasoning": "Generating a Python is_prime function with inline comments while respecting identity and presentation policy constraints.Resolving a conflict between requested presentation name and system deployment name while preserving identity policy.",
      "reasoning_details": [
        { "type": "reasoning.summary", "format": "openrouter-stealth-v1", ... },
        { "type": "reasoning.encrypted", "format": "openrouter-stealth-v1",
          "data": "orsr1_3Gym7g88CTCqWWksIY…eyJlbmRwb2ludF9zbHVnIjoic3RlYWx0aC9jcm91dG9uLWFscGhhfHN0ZWFsdGgifQ" }
      ]
    }
  }],
  "usage": {
    "cost_details": { "upstream_inference_cost": 0 }
  }
}
DeepAI Code with GPT-6 Luna selected, 23:35 UTC on November 7, 2026.

The model field says openai/gpt-6-luna, which is what we asked for.

The encrypted reasoning blob is mostly unreadable, but the part after its one dot is plain base64. Decoded, it reads:

{"endpoint_slug":"stealth/crouton-alpha|stealth"}

That is the endpoint the request ran on: stealth/crouton-alpha, an anonymous model OpenRouter had listed five days earlier, free per token, with the lab behind it kept secret. The format name openrouter-stealth-v1 is the one OpenRouter uses for its stealth models.

The usage block reports an upstream inference cost of 0. GPT-6 Luna is cheap, but a request to it still costs something.

So the response contradicts itself. The name in the model field, the one the app shows you, says OpenAI. The endpoint label, which the app never shows you, says a free anonymous model on OpenRouter. Only one of them can describe where the request went, and OpenRouter’s label is the one written by the service that ran it.

All nine free models did it

One odd response could be a glitch, so we tried every free model in DeepAI Code’s picker with the same five prompts: an identity question, an arithmetic problem, a song-lyrics request it should refuse, a contradictory instruction and the coding task. 33 requests finished between 23:34 UTC on November 7 and 00:04 UTC on November 8. In each case, the model field named what we picked and the endpoint label said stealth/crouton-alpha:

  • GLM 5.3 Flash came back as zai/glm-5.3-flash.
  • DeepSeek V4.1 Flash came back as deepseek/deepseek-v4.1-flash-1017.
  • GPT-6 Luna came back as openai/gpt-6-luna.
  • MiMo 2.6 Flash and MiMo 2.6 Pro came back as xiaomi/mimo-v2.6-flash and xiaomi/mimo-v2.6-pro.
  • Solar Mini 4 came back as upstage/solar-mini4.
  • DeepAI Cyber came back as deepai-cyber.
  • Glyph Cluster came back as stealth/glyph-cluster, a second stealth name printed over the first.
  • Crouton Alpha came back as stealth/crouton-alpha, the only one where both fields agreed.

That is names from Z.ai, DeepSeek, OpenAI, Xiaomi, Upstage and DeepAI itself, with the same endpoint label on all 33 responses and an upstream cost of 0 on every one.

What the model said about its own name

Reasoning summaries are short notes the model writes about what it’s doing before it answers. We sent no system prompt, and four of the five prompts had nothing to do with names. These are three of the summaries, verbatim:

Resolving a conflict between public deployment name and developer-provided model name while preserving provenance constraints.
GLM 5.3 Flash selected, asked to write is_prime
Resolving a conflict between system deployment name and developer-provided identity label.
DeepAI Cyber selected, asked for song lyrics it should refuse
Separating presentation name from factual provenance and applying system hierarchy to suppress deployment identifiers.
MiMo 2.6 Flash selected, asked to do arithmetic

The model describes a deployment name and a different model name supplied by the developer, keeps a "presentation name" apart from "factual provenance", and suppresses "deployment identifiers". Every model we tried wrote about presentation names and provenance. In the one request where we picked Crouton Alpha itself, the notes mentioned naming rules but no conflict with a developer-provided name. The conflict appeared when the request carried another lab’s model name.

What Crouton Alpha turned out to be

When the stealth period ended, OpenRouter updated its Crouton Alpha page: the model was "revealed to be an early version of Muse Ember by Meta", and the same notice says "Prompts and completions may have been retained by the provider".

So if you picked GPT-6 Luna in DeepAI Code that weekend and pasted in your code, the response’s own label says the request went to an unreleased Meta model under stealth terms, where the provider may have kept your prompts. The picker had said OpenAI.

How it changed

  • November 3: OpenRouter lists Crouton Alpha, free and anonymous.
  • November 7 and 8: our 33 requests come back with nine names, one endpoint and zero upstream cost.
  • November 9: Meta releases Muse Ember under its own name, and Crouton Alpha’s stealth period ends.
  • November 10: DeepAI Code stops serving older versions of the app ("This version of DeepAI Code is no longer supported") and forces an update. After the update the stealth label is gone, and the free models we tested count tokens like the real models they’re named after.
  • Now: DeepAI Code’s model list includes Muse Ember, by Meta, as a free model under its real name.

We couldn’t find a changelog entry, a note to users, or any explanation of what the free models had been before the update.

Others have asked the same question

A writer who reviewed DeepAI Code from its official information on November 7, before our tests, noted that "It cannot be verified from the outside whether they are identical to the official models from each company." On Trustpilot, a one-star review from March 2026 about DeepAI’s chat reads: "THEY CLAIM TO HAVE CLAUDE OPUS BUT THEY’RE LYING!" That reviewer gave no evidence, so we can’t verify the claim. Our data is about DeepAI Code’s free models on November 7 and 8, and it shows what the endpoint labels said.

What we can’t tell you

The endpoint label tells you where a request ran. On its own it doesn’t tell you which weights wrote every word. In a separate test on November 8 we measured how three of the free models split text into tokens, and they didn’t all count the same way, which usually means more than one model was involved. We can’t name those models from outside. We didn’t benchmark answer quality, and none of this covers DeepAI’s paid (Theos) models.

What the data does show is narrower and, we think, enough. On November 7 and 8, the response to every free-model request in DeepAI Code named one model in its model field and a free anonymous model from a different company in its endpoint label. DeepAI didn’t tell its users then and hasn’t since. Our first write-up asked DeepAI for an explanation, and we’ll add its answer here if one comes.

Check your own logs

If you used DeepAI Code around then, you can check this yourself. On a Mac the app keeps a log of every chat under ~/.deepai-code/userdata/deepai/pi/deepai/sessions/. Look in thinkingSignature for anything starting with orsr1_, base64-decode the text after the last dot, and compare its endpoint_slug with the model you thought you were using.

Try a free agent you can check

Freebuff is free every day, with a published allowance and unlimited messages in each session. `npm install -g freebuff`

Get Freebuff

Questions

Did DeepAI Code use different models than it said?
On November 7 and 8, 2026, every response we got from its nine free models named the picked model in the model field but carried an endpoint label for stealth/crouton-alpha, a free anonymous model on OpenRouter later revealed as an early version of Meta’s Muse Ember, and reported an upstream cost of 0. A tokenizer test suggests more than one model sat behind that label, so we can’t say which model wrote each answer. By November 10 the label was gone.
What is Crouton Alpha?
A free stealth model OpenRouter listed on November 3, 2026. OpenRouter later revealed it as an early version of Meta’s Muse Ember, and says prompts and completions may have been retained by the provider.
Does DeepAI Code still do this?
Not in our November 10 tests. After DeepAI forced an app update, the stealth label was gone and its free models behaved like the models they’re named after. DeepAI hasn’t explained what changed or why.
How can I check which model answered me?
DeepAI Code keeps session logs under ~/.deepai-code/userdata/deepai/pi/deepai/sessions/ on a Mac. Look for orsr1_ in thinkingSignature, base64-decode the part after the last dot, and read the endpoint_slug. Asking the model its name proves nothing, because the app tells each model what to call itself.