The Beast Max Challenge: Beat It, and the Request Is on Us

We think Beast Max at high effort gives the strongest answer you can buy for a hard problem. Until October 15 we are putting a price on that. Run any prompt through Beast Max at vhigh or above, then run the same prompt through the flagship model of any major lab. If their answer is objectively better than ours, we refund the full cost of the Beast Max request to the card or account you paid with.

The offer

Model: beast-max with reasoning_effort set to vhigh, xhigh, or yolo. In the Chat UI that is the VHigh, XHigh, or Yolo button next to Effort.
Comparison: the same prompt and attachments, one turn, against the current flagship model from OpenAI, Anthropic, Google, or xAI at its highest reasoning setting.
Outcome: if their answer is objectively better than ours, the full cost of the Beast Max request is refunded to your original payment method rather than credited to your account.
Window: September 24 through October 15, 2026, and we can stop it at any time. We will do so if it turns out we were wrong about our own model, since we would like to still exist in November.

What “objectively better” means

We mean something a third party could check without taste entering into it. A claim qualifies when the other model’s answer is right and ours is wrong in one of these ways.

Code. Their output passes a test suite, compiles, or produces the correct result on a defined input, and ours does not.
Math and logic. Their proof or calculation holds and ours contains an error.
Facts. They got a fact right that we got wrong, with a citable source.
Hard requirements. The prompt stated a constraint such as a format or a length limit, and they met it while we missed it.

Tone, length, formatting preference, an answer that simply reads better, a faster response, and a refusal on either side do not qualify. Speed in particular is not the contest. Beast Max deliberates before it answers, and we have never claimed otherwise.

How to claim

A claim needs the Beast Max conversation itself, the request id of the turn you are disputing, and the other model’s full output with the settings you used. Email all of it to [email protected] together with a sentence or two on what was wrong with ours, what was right with theirs, and how we can check.

We ask for the whole conversation because we keep request content only for a short monitoring window, typically a few days, and after that it is gone. The id tells us that a request happened and what it cost, but not what was said. The conversation you send is the record we judge against, so a claim without it cannot be judged.

From the Chat UI

Pick Beast Max in the Model selector and set Effort to VHigh, XHigh, or Yolo. Send your prompt.
On the Beast Max reply, use the Copy request details for support button in the message action row (the circled i icon). It copies a block with the request id, the model, and the effort level. Paste that block into your email.
On the last message of the conversation, use the Export conversation as Markdown button (the download arrow). Attach the .md file. The export carries the messages, token counts, and cost, but not the request id, which is why step 2 exists.

From the API

Every response carries its id in the X-Request-ID header, and Chat Completions responses also return it as beast_request_id in the body. Send the id, the exact request body you sent, and the response you received.

curl --max-time 7200 -D headers.txt https://api.beastlab.ai/v1/chat/completions \
  -H "Content-Type: application/json" \
  -H "Authorization: Bearer $BEASTLAB_API_KEY" \
  -d '{
    "model": "beast-max",
    "messages": [{"role": "user", "content": "Your hardest problem goes here."}],
    "reasoning_effort": "vhigh"
  }'

From a coding agent

If Beast Max was running inside a coding agent, send the agent’s exported session log or transcript instead of a Chat UI export. Most agents can export a session, and the rest keep the transcript on disk. Include the request ids where your tooling exposes them. Where it does not, we match on timing and content. The comparison run is the same task, same starting state, on the other lab’s flagship at its highest setting.

We reproduce the check and reply within five business days. Approved claims are refunded to your original payment method for what the request cost. Declined claims get a reason.

Any effort at vhigh or above qualifies. Start there and move to xhigh or yolo only where the problem earns it, since each level costs more and takes longer, as described in the API reference.

We will probably publish the losses

If a top-tier model beats Beast Max on a checkable task, that is exactly the case we want in front of our engineers, and, with your permission, in front of you as well, with the prompt, both answers and the criterion. We reserve the right to write these up with slightly less enthusiasm than this paragraph suggests, depending on how many there are. If there are none by October 15, expect a post about that with considerably more.

The fine print

The promotion runs from September 24, 2026 00:00 UTC through October 15, 2026 23:59 UTC, or until we end it, whichever comes first. We may end it early at any time by updating this post, and the time of that update is then the close of the window. The Beast Max request must be timestamped inside the window as so defined. Requests made after an early close are not eligible. Claims for requests made before the close are honored under the rules in force when the request was made.
Eligible requests are to beast-max only, with reasoning_effort of vhigh, xhigh, or yolo, sent top-level or as the nested reasoning.effort form, or the matching Effort setting in the Chat UI. Requests at lower effort, or to Beast Mini or Beast Nano, are not eligible.
The comparison must be a single turn on the current flagship model from OpenAI, Anthropic, Google, or xAI at its highest reasoning setting on the day, with the same prompt, attachments, and system prompt sent to Beast Max.
A claim must include the full conversation or session log, the request id of the disputed turn, and the other model’s full output and settings. Claims without the conversation cannot be judged and will be declined.
A request that never produced an answer does not qualify. If you got an error, a timeout, an empty or cut-off response, or output that is plainly broken by a fault or outage on our side, that is a service problem and support will sort it out. It is not a claim under this challenge, because there has to be a completed Beast Max answer to judge. Responses cut short by a client-side limit such as max_tokens are not eligible either.
One claim per prompt per account. Refunds are capped at US$500 per account over the promotion.
Refunds are paid to the payment method used for the purchase that funded the request rather than issued as account credit, and cover the metered cost of the eligible request as billed. Where the payment provider cannot return funds to that method, we agree an equivalent method with you. Refunds are issued within 10 business days of approval, although the provider may take longer to show it. Requests paid from free or promotional credit involved no payment and are not refundable.
Claims must be submitted within 7 days after the window closes, including an early close. We reply within five business days. Our determination of whether an answer is objectively better is made in good faith and is final.
Prompts engineered so that no model can answer them correctly, prompts designed to elicit a refusal, and prompts that violate our Terms are excluded.
The challenge is about hard problems. Trick questions, riddles, tokenizer traps, counting-letters puzzles, and other gotcha prompts that test a quirk rather than reasoning do not qualify. If a task would not need vhigh to begin with, it is not a claim.
For eligible requests only, this promotion supersedes section 6.5 of the Terms.

Hard problems welcome. See the full Beast Max specs and pricing, or start on the Portal.

Run Beast Max at vhigh on the Portal