Models

MiniMax says its secret model handled 200 trillion tokens. The host that ran it counted 56.7 trillion.

MiniMax revealed M3.1 Flash on October 19 as the stealth model Space Bunny. The price is $0.15 per million input tokens. The usage figure in the launch post is about 3.5 times what OpenRouter reported.

·4 min read
Illustration of a bar chart with two bars of very different height, for the story on MiniMax M3.1 Flash and its stealth-run token counts

MiniMax’s launch post for M3.1 Flash says the model, while it was running anonymously as “Space Bunny,” “processed more than 200 trillion tokens.” OpenRouter, the platform that hosted the stealth run, says 56.7 trillion.

Both are quoted on October 19, the day the China-based lab took the codename off. They can’t both describe the same thing, and neither company explains the gap.

Where the two figures could come from

OpenRouter’s number is scoped to OpenRouter. It says Space Bunny was its top model for 10 days in a row and the busiest stealth model it has hosted, and that it left the platform on October 5 and came back on October 19 under its real name. MiniMax’s wording is wider: “while operating anonymously.” Read literally, that could include traffic MiniMax served itself or through other partners, which OpenRouter wouldn’t see.

That’s a guess. Neither source says it. What I can say is that the 200 trillion figure comes from the company selling the model, with no breakdown, and the 56.7 trillion figure comes from a third party with a clear scope. If I had to repeat one of them, it’d be the second.

What else MiniMax is claiming

The rest of the launch post is the usual set of company-reported results. On DeepSWE v1.1, a coding benchmark, it scores 73.7%. It gets 61.7 on OSWorld V2 (computer use), 82% on CyberGym (finding and confirming vulnerabilities) and 76.2 on BankerToolBench. MiniMax says it ranks first among Chinese models on OSWorld V2 and on Agents’ Last Exam. Nobody outside the company has re-run any of this yet, and the post gives no comparison table I could check against another lab’s published numbers.

The speed and cost claims are more concrete. MiniMax says online inference is nearly three times faster and agentic tasks cost 70% less than on the previous model, M3. Output speeds, it says, reach 200+ tokens a second, with a median above 170. The list price is $0.15 per million input tokens, $0.60 per million output tokens and $0.015 for cached input.

OpenRouter’s reveal adds the stealth-period changes: time to first token fell from 1.95 to 1.22 seconds, throughput rose from 70 to 100 tokens a second, and errors fell from 1.4% to 0.7%. Those are tidy numbers about a model that was being tuned live on user traffic, and they come relayed through a trade article, not from OpenRouter’s page itself, which I couldn’t open.

Weights: two dates, one of them vague

The launch post says the model’s weights “will be released later.” The Unite.AI write-up says MiniMax’s launch post on X gave a date of October 23. I couldn’t open the X post, so treat the date as likely but unconfirmed. A Pro model is promised after that, with no date at all.

For a US team, the open-weights question matters more than the leaderboard. A model you can download can be run on your own servers and audited. One you can only rent through a Chinese company’s API, or through a US reseller, can’t. MiniMax has open-sourced models before (Unite.AI has covered its M2.7 release), but I’d wait for the files. Cheap prices like this also keep pressure on US labs, as the GPT-6 price cuts showed.

What I think

The price is the real story, and it’s the part that’s easiest to check: $0.15 in and $0.60 out is cheap for something claiming coding scores in the 70s. It also sits right next to DeepSeek’s V4.1-Flash-1017 from two days earlier, which also claimed a strong DeepSWE result under an open license. Two Chinese labs in one weekend are now both saying “frontier-level coding for pocket change.”

I’d treat the 200 trillion as marketing until MiniMax explains it. A stealth model that wins by usage is a clever way to get free testing. It’s less clever to publish a headline number that your own hosting partner’s number undercuts by a factor of 3.5. Independent runs of DeepSWE and OSWorld, and the actual weights on October 23 if they appear, will say more than either token count.

Sources

  1. MiniMax M3.1 Flash: Redefining Agent Performance and Cost (MiniMax, October 19, 2026)
  2. MiniMax Launches M3.1 Flash, Revealing Space Bunny Stealth Model (Unite.AI, October 19, 2026; relays OpenRouter’s reveal post and MiniMax’s X post)
  3. MiniMax says M3.1 Flash makes online inference nearly three times faster and cuts costs 70% (RuntimeWire, October 19, 2026)