Crouton Alpha: Benchmarks, API Access and Who Made It

Crouton Alpha is a stealth reasoning model for agent sub-tasks, coding and document analysis. An anonymous provider released it on November 3, 2026. Its developer has not been confirmed.

You can run Crouton Alpha through AI/ML API with the model ID stealth/crouton-alpha. It has a 1M-token context and is free during the current preview, according to our model listing, checked on November 5, 2026.

In SMF Works' Cold Iron run, Crouton Alpha passed 140 of 157 tests, two fewer than Claude Fable 5.1. In our test, it passed all nine coding, tool and log tasks. Across the six short coding tasks, it had the lowest median completion time of the three models we compared, ahead of GLM 5.3 Flash and MiniMax M3.1 Flash.

We ran our tests on November 4, 2026. We rechecked public model listings, API documentation and listed prices on November 5.

What is the Crouton Alpha stealth model?

A stealth model is available under a temporary name while its provider remains anonymous.

OpenRouter's listing describes Crouton Alpha as a lightweight model for repeated agent sub-tasks, tool calls, coding steps and document analysis. OpenRouter routes requests to the model; an anonymous third party develops and operates it.

Developers can pair it with a frontier reasoning model. Crouton Alpha then does the small, clearly defined steps that turn a plan into action.

Crouton Alpha specs

The AI/ML API reference documents image, video and PDF input, tool calls and JSON output. Our image and video requests worked; the two audio formats we tried returned errors.

How good is Crouton Alpha on benchmarks?

The anonymous provider has not published an official benchmark report. The results below come from two independent testers: SMF Works, with Cold Iron, and MindStudio.

Cold Iron has 157 pass or fail tests in seven areas, from coding to tool calls. SMF Works ran Crouton Alpha on November 3 and shared the result files on GitHub.

SMF Works used minimal reasoning effort. The API reported reasoning tokens on 26 of the 157 tests; a zero count on the other requests does not establish that the model did no reasoning.

None of the 11 math misses hit the token limit. The whole run took 371.7 seconds, and the median answer came back in 2.12 seconds.

MindStudio gave Crouton Alpha three small coding tasks in Hermes Agent, which sent each task to its own sub-agent. After the model said it was done, MindStudio reran the unit tests: all 20 passed.

In a second test, MindStudio gave it a hand-drawn sketch of a kebab with messy handwritten notes, such as "slices fall here". In about three minutes, Crouton Alpha returned a web page where the kebab spins and slices fall on the side the note pointed to. The knife didn't line up with the kebab, and some shapes were rough.

Crouton Alpha vs. other models

In SMF Works' November 3 Cold Iron post, Crouton Alpha ranked 10th among the 26 models on that board. The table below reproduces selected published places. These are single-run scores with different reasoning settings, so they do not establish a general ranking of the models.

Grok 4.6 and Claude Fable 5.1 ran with reasoning turned off, and Gemini 4 Argon ran at its default setting. Crouton Alpha ran at minimal, which still allows reasoning; MiniMax M3.1 Flash ran at low. SMF Works reports that every higher-scoring run took at least 900 seconds. The scores came from OpenRouter runs; the prices above are AI/ML API's listed rates, checked on November 5.

How did Crouton Alpha do in our test?

All three models passed the six coding tasks, two tool tasks and one log task. Crouton Alpha had the lowest median coding time in this sample: 3.8 seconds, against 32.6 for GLM 5.3 Flash and 4.7 for MiniMax M3.1 Flash.

How we tested

We ran the test through AI/ML API on November 4, 2026. We compared Crouton Alpha with GLM 5.3 Flash and MiniMax M3.1 Flash, the released versions of the two stealth models we covered before. In this comparison, each model ran at its default reasoning effort.

Crouton Alpha read the chart in our image and correctly described the video of a flower bud opening in a time-lapse. It couldn't answer the audio question, because AI/ML API rejected both audio formats we tried. Our tokenizer test is in the section about who made Crouton Alpha.

Results by task type

For this log, AI/ML API reported 25 to 28% fewer input tokens for Crouton Alpha than for the other two models.

At minimal effort, the first output from a 300-word story arrived after 1.5 to 1.8 seconds. Output speed was 114 to 129 tokens per second. At the default effort, the first output arrived after 8.5 to 9.6 seconds. AI/ML API reported no reasoning tokens on 29 of our 36 coding runs and at most 125 on the rest. These counts do not show how long the model spent reasoning.

Which reasoning effort to pick

Start with minimal for tasks like the six short coding problems we tested. It solved all six, matching the default effort, with a median of 2.1 seconds per task versus 3.8 seconds.

Our tasks were short, so this test can't show where a higher effort level helps.

Who made Crouton Alpha?

The developer remains unknown. Some settings resemble Meta's Muse models, while reported token counts differ. These comparisons do not identify the developer.

How Crouton Alpha compares with Meta Muse

We compared the OpenRouter catalog on November 5. The token-count experiment below comes from our November 4 tests.

Our tests did not identify the provider. Some settings resemble Meta's Muse models, while reported token counts differed from both Muse models in our 23-string test. Request formatting, service-added tokens and provider accounting can affect those counts, so the mismatch does not establish a new tokenizer or rule out a developer.

Two previous stealth models we covered later received public names. Their reveal timelines do not predict when Crouton Alpha might receive one.

Ox Alpha appeared on August 20, 2026. Six days later, Zhipu AI released it as GLM 5.3 Flash. Its tokenizer had already pointed to Zhipu AI.

Space Bunny Alpha appeared on September 23. On October 19, MiniMax said it was an early version of MiniMax M3.1 Flash. Our MiniMax M3.1 Flash review has the details.

Union Alpha, listed at 12th in the same Cold Iron post, was later revealed as Pareto from Unbiased.

How to use the Crouton Alpha API

You can try Crouton Alpha in the AI/ML API Playground or call it from your code.

Option 1: Test Crouton Alpha in the Playground

Open Crouton Alpha in the AI/ML API Playground, sign in and send a prompt. For a request with explicit reasoning effort or a repeatable test, use the API example below.

Option 2: Call Crouton Alpha through the API

1. Create an API key

Open the API keys page, click Create API Key and make sure the key is enabled. Store the key in an environment variable. Don't put it in browser code or commit it to a repository.

On macOS or Linux:

export AIMLAPI_KEY="YOUR_AIMLAPI_KEY"

On Windows PowerShell:

$env:AIMLAPI_KEY="YOUR_AIMLAPI_KEY"

2. Send a request

Crouton Alpha uses the OpenAI-compatible Chat Completions endpoint:

POST https://api.aimlapi.com/v1/chat/completions

For the full schema, see the Crouton Alpha API reference.

Every request needs two fields:

On macOS or Linux, run this command in Bash or Zsh:

curl --fail-with-body --silent --show-error --request POST \
  --url "https://api.aimlapi.com/v1/chat/completions" \
  --header "Authorization: Bearer $AIMLAPI_KEY" \
  --header "Content-Type: application/json" \
  --data '{
    "model": "stealth/crouton-alpha",
    "messages": [
      {
        "role": "user",
        "content": "Return JSON with the invoice number and the total: Invoice INV-2041, total 1,240.00 USD."
      }
    ],
    "reasoning_effort": "minimal",
    "response_format": {"type": "json_object"},
    "max_completion_tokens": 2000
  }'

On Windows, use this PowerShell request:

$body = @'
{
  "model": "stealth/crouton-alpha",
  "messages": [
    {
      "role": "user",
      "content": "Return JSON with the invoice number and the total: Invoice INV-2041, total 1,240.00 USD."
    }
  ],
  "reasoning_effort": "minimal",
  "response_format": {
    "type": "json_object"
  },
  "max_completion_tokens": 2000
}
'@
$headers = @{ Authorization = "Bearer $env:AIMLAPI_KEY" }
$response = Invoke-RestMethod -Method Post -Uri "https://api.aimlapi.com/v1/chat/completions" -Headers $headers -ContentType "application/json" -Body $body
$response.choices[0].message.content

Both examples use reasoning_effort: "minimal" for this short extraction task. If you omit it, the catalog default is max.

3. Check the response

The answer is in choices[0].message.content. Check the HTTP status and any API error before reading it. Then check choices[0].finish_reason: stop means generation finished, and length means it reached the token budget. Check that the content is nonempty before parsing it as JSON, then validate the fields and values your application needs. When you supply tools, tool_calls is a separate finish reason that asks your application to execute the requested functions.

Crouton Alpha can spend tokens on hidden reasoning before it answers. In our check on AI/ML API, a request with a 64-token limit at the default reasoning effort came back empty, with finish_reason set to length. With reasoning_effort set to minimal, the same request returned the answer. Leave enough room in max_completion_tokens.

FAQ

Is Crouton Alpha free?

AI/ML API lists Crouton Alpha as free during the current stealth preview, checked on November 5, 2026. Preview pricing can change. Earlier stealth previews have ended: the former Ox Alpha is now GLM 5.3 Flash, listed at $0.15 per million input tokens and $0.50 per million output tokens on AI/ML API.

Does Crouton Alpha keep my data?

OpenRouter's stealth-model listing says the underlying provider may retain prompts and completions but does not use them for training. That statement does not describe every intermediary's data handling. Check the terms of the API service you use before sending private code, secrets or personal data.

Is Crouton Alpha made by Meta?

The developer has not been confirmed. OpenRouter lists the same six reasoning effort levels for Crouton Alpha, Muse Wonder and Muse Spark 1.3. In our 23-string test, its reported token counts matched neither Muse model. Neither observation establishes who made it.

Is Crouton Alpha better than Claude or GPT for coding?

There is no evidence for that yet. Cold Iron's coding tasks are short: Crouton Alpha's longest coding answer there was 566 tokens. Its perfect coding score says little about work on large codebases. For bigger coding tasks, see our list of the best LLMs for coding.

Is Crouton Alpha open source?

No public model weights or open-source license have been announced for Crouton Alpha. You can access the preview through hosted services such as AI/ML API.

Key takeaways

• Math accounted for 11 of its 17 Cold Iron failures: it solved 19 of 30 tasks.
• Across our six short coding tasks, it had the lowest median time of the three models: 3.8 seconds at default effort and 2.1 at minimal.
• For our test log, AI/ML API reported 25 to 28% fewer input tokens than for GLM 5.3 Flash and MiniMax M3.1 Flash.
• Shared settings and differing reported token counts do not identify Crouton Alpha's provider.
• Call it through AI/ML API with the model ID stealth/crouton-alpha.