Space Bunny has a name
On October 19, 2026, MiniMax released MiniMax M3.1 Flash and said Space Bunny, the anonymous model listed on OpenRouter since September 23 as Space Bunny Alpha, was an early version of it. MiniMax’s pricing page lists $0.15 per million input tokens, $0.60 per million output tokens, and $0.015 for cached input. The model guide gives MiniMax-M3.1-Flash a 1,000,000 token context window, with text, image, and video in. Calls go through the MiniMax API, the M Plan, and the MiniMax Code desktop app, and the announcement says the weights will be released later.
The short version: the token got cheaper, but this model always thinks, and the thinking is what you end up paying for. I read MiniMax’s announcement, the API docs, the pricing page, OpenRouter’s stealth page and Plotly’s write-up. I did not run the model myself, so every score below is either MiniMax’s own claim or Plotly’s single test.
Nearly four weeks without a name
OpenRouter’s stealth page describes Space Bunny Alpha as an anonymous model with fast inference, strong coding, native multimodal input, adjustable reasoning effort, a 1 million token context window, text, image, and video in, and text out. The listing date is September 23, 2026. The page now says the model was revealed as MiniMax M3.1 Flash. The free stealth period has ended. Prompts and completions may have been retained by the provider, and they are not used for training.
Read that wording slowly if you pasted client code into it during the free month. Retained is possible, training is ruled out, and neither word means deleted.
MiniMax says the anonymous run processed more than 200 trillion tokens. TokenPost, writing on October 19, reports that the preview appeared on OpenRouter on September 23 and that by September 24 it held the top spot in OpenCode’s daily token-usage ranking. OpenCode’s data page lists the named model second by tokens across OpenCode last week, and at 16 percent of tokens over the past two months.
Why launch anonymously? My guess, and it is only a guess: a model with no name gets judged on its output, not on the logo. What MiniMax actually published is the token count, and 200 trillion is a lot of free testing.
What MiniMax M3.1 Flash claims on paper
According to MiniMax’s announcement, the release is “a generational leap in agentic tasks, coding, Computer Use and cybersecurity,” online inference is “nearly three times faster,” and overall usage cost on agentic tasks is 70 percent lower than its predecessor. The speed claim names no baseline, and MiniMax doesn’t say which model “predecessor” means. So I won’t guess, and I’ll compare against MiniMax-M3’s price card further down instead.
Benchmarks from the same post, all MiniMax’s claims:
- DeepSWE v1.1 at 73.7 percent
- OSWorld V2 at 61.7
- Agents’ Last Exam at 29.6, and MiniMax says first among Chinese models on both OSWorld V2 and ALE
- BankerToolBench at 76.2
- CyberGym at 82 percent, for vulnerability discovery and validation
- AutomationBench-AA, “among the top performers” on enterprise SaaS workflows in sales, marketing, operations, customer service, finance, and HR
MiniMax adds that its internal security experts find the model strong in code auditing, vulnerability patching, penetration testing, malware analysis, and reverse engineering. No method or numbers come with that one.
Speed, same announcement: “more aggressive sparsity,” output that can pass 200 tokens per second, and a median (P50) above 170 tokens per second.
Nobody outside MiniMax has reproduced these numbers yet. The one outside test I found is Plotly’s, and it gets its own section below.
The model guide calls MiniMax-M3.1-Flash a “frontier multimodal coding model” with tunable thinking depth, built for agentic reasoning, tool use, coding, and structured task execution. (I wrote up what the frontier label actually means yesterday.) The same guide also calls it “an ultra-efficient open-weight model for everyday work,” which doesn’t square with the announcement’s “later” on weights. More on that below.
MiniMax-M3 remains on the guide, also with a 1 million token context window.
MiniMax’s OpenAI SDK page lists the file rules. Images: JPEG, PNG, GIF, or WEBP, up to 10 MB. Videos: MP4, AVI, MOV, or MKV. Default sampling is 1 frame per second, allowed from 0.2 to 5. URL or base64 video tops out at 50 MB. Files API video tops out at 512 MB. Audio input is not supported.
What does it cost?
US dollars per 1 million tokens. Flash and M3 come from MiniMax’s pay-as-you-go page. DeepSeek comes from DeepSeek’s pricing page, where deepseek-flash is DeepSeek-V4.1-Flash-1017, with a 1 million token context window.
| Model | Input | Output | Cached input |
|---|---|---|---|
| MiniMax M3.1 Flash | $0.15 | $0.60 | $0.015 |
| MiniMax M3 (standard, ≤512k input) | $0.30 | $1.20 | $0.06 |
| DeepSeek V4.1 Flash (off-peak / peak) | $0.15 / $0.30 | $0.60 / $1.20 | $0.003 / $0.006 |
On input and output, MiniMax M3.1 Flash is half of MiniMax-M3’s standard rate at or under 512k of input. Cached input on Flash is $0.015, a tenth of Flash’s $0.15 input price and a quarter of M3’s $0.06 cache price at that tier.
Above 512k of input, M3 is priced at $0.60, $2.40, and $0.12. The page also lists a priority tier at 1.5x for M3. Flash’s published line stays flat at $0.15 / $0.60 / $0.015.
MiniMax’s 70 percent figure is about usage cost on agentic tasks, against that unnamed predecessor. A per-token sticker tells you very little until you know how many tokens, calls and cache hits one task burns. That is the next section.
Off-peak, DeepSeek matches Flash on input and output at $0.15 and $0.60. Peak hours on DeepSeek’s page are 01:00 to 04:00 UTC and 06:00 to 10:00 UTC, Monday to Friday, and in those hours input and output double to $0.30 and $1.20. Cache hits stay lower than Flash in both windows, $0.003 off-peak and $0.006 peak, against $0.015. One difference the table can’t show: DeepSeek-V4.1-Flash-1017 already shipped as open weights (it was one of three open-weight releases in the October 19 digest), while Flash weights are still a promise. The $0.15 / $0.60 pair also lands right on the cheap-tier floor I wrote about in the Haiku 5.5 and GPT-6 Luna note, with one extra rule on Flash: thinking stays on.
Cheap tokens, long thinking
MiniMax’s chat completion reference says M3.1 Flash always thinks. Setting thinking to disabled returns HTTP 400. The trace comes back on its own, in reasoning_content. Effort is called reasoning_effort there and effort in the model guide. The accepted values are low, medium, high, xhigh, and max. The default is max.
MiniMax-M3, on the same page, thinks adaptively, and disabled really does skip the thinking. So the trade is plain. Flash has the cheaper token and no way to skip the thinking. M3 lets you skip it, at twice the price. And a request body copied from an M3 integration with thinking disabled will bounce off Flash with HTTP 400.
Plotly’s October 20 write-up is the one independent test I found. Plotly ran the same 43 data-analytics questions against Space Bunny in September and against MiniMax M3.1 Flash in October. The write-up doesn’t say which effort level was used, so I won’t either.
September, then October:
- Score: 38 of 43, then 41 of 43
- Hidden patterns: 2 of 5, then 5 of 5
- Model calls per question: 15, then 45
- Tokens read per question: 0.7 million, then 4.2 million
- Tokens written per question: 8,500, then 32,000
- Median time per question: 36 seconds, then 3 minutes
- Full run at MiniMax list prices: $0.82, then $4.09
Look at the middle rows. Calls tripled, written tokens nearly quadrupled, tokens read went up six times. The released model works harder, and in Plotly’s reading that’s where the extra hidden patterns came from. Plotly also says 98 percent of the input was cached, which is the only reason a 4.2 million token question stays cheap. OpenRouter was serving M3.1 Flash for free for now, so Plotly priced the October run at list price and put it at about 10 cents a question.
On score, 41 of 43 ties Claude Opus 5.5, Gemini 4 Argon, Muse Wonder, and GPT-6 Sol. Plotly calls that the best score they have seen from a Chinese lab. Next on their list: GLM 5.3 Flash and Qwen 3.8 27B, both at 38. September’s Space Bunny, at 38 of 43, was “as well as any open-weight model we’d tested.” October added three correct answers and all five hidden patterns.
Plotly puts this 41 of 43 at about a third of the cost of Claude Opus 5.5 and Gemini 4 Argon for the same score. GPT-6 Sol scored 41 for $2.61, and GPT-6.1 Sol scored 43 of 43 for that same $2.61. One suite, one set of questions: a data point, not a ranking.
The clock is the other bill. Three minutes median is the slowest mark on Plotly’s leaderboard. They observed about 150 tokens per second per call. MiniMax cites a P50 above 170 and peaks above 200. Plotly’s check against that median: even at 170 tokens per second, writing 32,000 tokens takes about three minutes. Their judgment: fine for analysis that runs in the background, and slow for someone waiting on an answer.
Effort is the dial I would budget. Default is max, disabled is rejected, and I have no measurement of low against max. Plotly’s October model still called more and wrote more than the September preview on the same questions. Background work can live with a three-minute median. A person waiting on a chat answer will notice the wait first.
One more pricing wrinkle. OpenRouter’s stealth page says the free stealth period has ended, and Plotly wrote on October 20 that the named model was free on OpenRouter “for now”. Whatever you see on a router this week, I’d budget at MiniMax’s list price.
What changes if you were on M2.5 or M2.7?
MiniMax’s model list gives the M2.5 and M2.7 families two dates. From October 19, 2026, new users cannot start on them. On November 18, 2026, they are retired and API requests stop working.
Pay-as-you-go customers who already had access keep it until November 18. M Plan and Token Plan users are redirected. From October 19, requests to M2.5 or M2.7 route automatically to MiniMax M3.1 Flash. Old model names keep working. Plan allowances stay unchanged.
A plan user can change models without changing a line of code. Outputs, latency, and tokens per task can still shift, because the model behind the old name now always thinks, and thinking disabled returns HTTP 400. The allowance figure stays. How many tasks it covers depends on tokens per task. I would re-run evals on the work I actually ship, and I would treat old M2.7 logs as a different model.
MiniMax-M3 remains available, including the switch that skips thinking, at M3’s higher rates.
Open weights, eventually
MiniMax’s announcement says “Model weights will be released later.” Pro, which “pushes the boundaries of intelligence,” will also arrive later. The pages I read give no date, no parameter count and no license.
The docs already use a stronger phrase. The model guide calls MiniMax M3.1 Flash “an ultra-efficient open-weight model,” the announcement says later, and OpenCode lists open weights as No. As of October 21, 2026, I couldn’t find downloadable weights or a license, so there is nothing to plan a self-hosted copy around yet.
There is also a Beijing angle. Under China’s October 16 draft, covered in my note on China AI export controls and model weights, already-published open weights are out of scope and new open-source releases get automatic licensing. If that draft becomes final before MiniMax ships the file, the release would go through that route. The parameter count is unpublished, so I won’t guess whether any threshold applies.
Before you switch
- Run the same tasks at two or three effort levels and log calls, tokens, and time. The default is max. Choose it on purpose.
- Watch the cached share of input. At $0.015 against $0.15, that share moves the input side of the bill and leaves output tokens and extra calls untouched.
- Check the pipe. MiniMax Code, the MiniMax API, and an OpenRouter route can share a brand and still differ.
- Mark November 18 if anything in production still says M2.5 or M2.7. Plan traffic is already redirected. Pay-as-you-go calls on those names stop that day.
- Wait for weights and a license before a self-hosting plan.
The rate card is the easy page to read. The invoice counts the thinking.
Sources
- MiniMax announcement, October 19, 2026
- MiniMax model invocation guide
- MiniMax create chat completion
- MiniMax OpenAI SDK page
- MiniMax model list and retirement notice
- MiniMax pay-as-you-go pricing
- OpenRouter stealth page for Space Bunny Alpha
- TokenPost, October 19, 2026
- OpenCode data page for MiniMax M3.1 Flash
- Plotly data analytics write-up, October 20, 2026
- DeepSeek API pricing
