MiniMax Reveals Space Bunny Was M3.1 Flash at $0.15 per Million Tokens With Weights Due October 23, but Its Own Table Shows DeepSeek Ahead on Coding
MiniMax launched M3.1 Flash on 19 Oct 2026 and confirmed it was the stealth Space Bunny model. It costs $0.15 input and $0.60 output per million tokens, with weights due 23 Oct. We read its benchmark table, price math and migration dates.
MiniMax has confirmed that Space Bunny, the anonymous model that MiniMax says topped global usage leaderboards while it ran unnamed, was an early build of its new MiniMax M3.1 Flash. The Shanghai lab launched the model on 19 October 2026 at a flat $0.15 per million input tokens and $0.60 per million output tokens.
In its launch post on X, MiniMax said the model weights “will arrive on Oct 23” and claimed M3.1 Flash runs at one third the cost of M3 at three times the speed. The official MiniMax blog puts the saving at 70% for agentic tasks.
The launch is real and the price is low. But we went through MiniMax’s own benchmark table, its pricing page and its cost charts, and the picture is more mixed than the “frontier-level intelligence” headline. Here is what is confirmed, what is not, and what developers on M2.5 or M2.7 need to do before 18 November.
What MiniMax actually announced
M3.1 Flash is a multimodal model that takes text, images and video and returns text. MiniMax’s model invocation docs list a 1,000,000 token context window and five thinking levels set through an effort parameter: low, medium, high, xhigh and max.
It is live now on the MiniMax API, in the M Plan subscription and in the MiniMax Code desktop app. OpenRouter has a free, rate limited listing at minimax/minimax-m3.1-flash:free that supports tool calling and JSON output, with up to 524,288 completion tokens per request.
MiniMax says Space Bunny processed “more than 200 trillion tokens” while it ran anonymously. OpenRouter’s Space Bunny Alpha page now states the stealth model “was developed and operated by MiniMax” and lists its release as 23 September 2026. The free stealth period has ended.
The price is half of M3 per token, not one third
MiniMax’s pricing page shows exactly how the new model compares with M3, which launched on 1 June. INR figures below use about ₹95.18 to the dollar.
| Per 1M tokens | M3.1 Flash | M3 (up to 512K input) | M3 (over 512K input) |
|---|---|---|---|
| Input | $0.15 (about ₹14.3) | $0.30 (about ₹28.6) | $0.60 (about ₹57.1) |
| Output | $0.60 (about ₹57.1) | $1.20 (about ₹114.2) | $2.40 (about ₹228.4) |
| Cached input read | $0.015 (about ₹1.4) | $0.06 (about ₹5.7) | $0.12 (about ₹11.4) |
On a token for token basis, M3.1 Flash is 50% cheaper than M3 for input and output, and 75% cheaper for cache reads. The 70% and “one third” figures only hold if the new model also uses fewer tokens per task, which is a MiniMax claim we could not test.
The bigger saving is on very long prompts. M3 doubles its rate once input passes 512K tokens, while M3.1 Flash stays flat. An 800K token prompt costs $0.12 in input on M3.1 Flash and $0.48 on M3, a 75% cut.
One catch: the docs say the default effort, when you leave it out, is max. MiniMax’s own Automation-Bench chart shows max effort scoring 62.8% against 52.8% at low, and max is also the setting that burns the most thinking tokens. If cost matters, set effort yourself.
What the benchmark table really shows
MiniMax published a 15 row comparison table with its launch post. It compares M3.1 Flash against two DeepSeek V4.1 Flash builds, Kimi K3, GLM-5.3, GPT-5.6 Sol and Claude Opus 5. All numbers are vendor reported.
By our count, M3.1 Flash has the top score outright on two rows: Automation-Bench v1.0.6 (62.2) and LOCA-Bench 256K (60.3). It ties Claude Opus 5 on BankerToolBench at 76.2. Opus 5 leads outright on eight rows, GPT-5.6 Sol on two, and DeepSeek V4.1 Flash 1017 and GLM-5.3 on one each.
Coding is the weak spot. On DeepSWE v1.1, M3.1 Flash scores 73.7, behind DeepSeek V4.1 Flash 1017 at 77.0 and even the older September DeepSeek V4.1 Flash build at 74.2. On NL2Repo-Bench it trails both DeepSeek builds as well.
| Benchmark | M3.1 Flash | Best in MiniMax’s table |
|---|---|---|
| Automation-Bench v1.0.6 | 62.2 | 62.2 (M3.1 Flash) |
| DeepSWE v1.1 | 73.7 | 77.0 (DeepSeek V4.1 Flash 1017) |
| OSWorld V2 | 61.7 | 70.6 (Claude Opus 5) |
| Terminal Bench 4.0 | 38.1 | 51.8 (Claude Opus 5) |
| CyberGym | 82.2 | 84.5 (GLM-5.3) |
| BabyVision (with tools) | 91.0 | 94.1 (Claude Opus 5) |
Three more details are worth knowing. First, the table compares against Opus 5 and GPT-5.6 Sol, not the newer Claude Opus 5.5 or GPT-6 models. Those newer models only appear on MiniMax’s cost charts, where Opus 5.5 still has the highest Agents’ Last Exam pass rate.
Second, MiniMax’s numbers do not fully agree with each other. The table lists Automation-Bench at 62.2, while the cost chart on the blog shows 62.8% at max effort. MiniMax has not explained the gap.
Third, the Automation-Bench chart prices DeepSeek at its peak rates of $0.30 input and $1.20 output. DeepSeek’s off-peak rates are $0.15 and $0.60, the same as M3.1 Flash. Outside DeepSeek’s peak hours, the per token price advantage in that chart disappears.
MiniMax’s blog also says M3.1 Flash ranks first among Chinese models on OSWorld V2 and Agents’ Last Exam. In its own table, DeepSeek V4.1 Flash 1017 has no score on either test, so that ranking rests partly on missing entries.
Space Bunny users should read the stealth terms
Plenty of developers used Space Bunny for free through OpenRouter and coding agents. TokenPost reported it was top of OpenCode’s daily token usage ranking by 24 September.
OpenRouter’s page says prompts and completions sent to the stealth model “may have been retained by the provider but are not used for training.” If you pasted client code or keys into Space Bunny, treat that data as sitting with MiniMax. The recent OpenCode Console breach is a reminder that keys travel further than you expect in agent tools.
The 200 trillion token figure is MiniMax’s own and we found no independent breakdown of it. Claims on X that Space Bunny held first place on OpenRouter for 10 days are not confirmed by MiniMax’s blog.
Confirmed vs unconfirmed
| Claim | Status |
|---|---|
| Space Bunny was an early M3.1 Flash build | Confirmed by MiniMax and OpenRouter |
| $0.15 input, $0.60 output, $0.015 cached, flat at any length | Confirmed on MiniMax’s pricing page |
| 1M context, five effort levels, default max | Confirmed in MiniMax docs |
| Weights on 23 October | Stated by MiniMax on X; the blog only says “later”; license not announced |
| 70% cheaper per task, 3x faster than M3 | Vendor claim, not independently tested |
| 200+ tokens per second output | Vendor claim |
| Benchmark scores | Vendor reported; no independent results found at publication |
| Top of OpenRouter for 10 days | Unconfirmed social media claim |
M2.5 and M2.7 users have a deadline
The launch comes with a retirement notice in MiniMax’s model FAQ. From 19 October, M2.5 and M2.7 are closed to new users, and on 18 November 2026 they are retired.
Pay as you go API users can keep calling them until 18 November, after which requests stop working. M Plan and Token Plan users have already been moved: since 19 October, requests that name M2.5 or M2.7 are routed to M3.1 Flash automatically, with no change to usage allowances.
That silent switch matters if you depend on fixed outputs, prompt formats or evals. Your model name has not changed, but the model behind it has.
What it means for Indian developers
At about ₹14.3 per million input tokens and ₹57.1 per million output tokens, M3.1 Flash matches DeepSeek’s off-peak rate and costs a fraction of Western frontier models, and you can call it through Anthropic or OpenAI compatible endpoints. The free OpenRouter listing is a low risk way to test it on your own tasks before paying.
MiniMax’s rate limits page lists 200 requests per minute and 10 million tokens per minute for M3.1 Flash and M3, which is enough for most startup workloads. Teams running batch jobs should still run their own cost test with effort set below max.
If the weights do land on 23 October, self hosting becomes possible for teams with data residency concerns, though no parameter count or hardware guide has been published yet. China’s draft export rules for large AI models exempt published open weights, so a public release should stay downloadable from India.
FAQ
Is MiniMax M3.1 Flash free?
There is a free, rate limited version on OpenRouter. On the MiniMax API it costs $0.15 per million input tokens and $0.60 per million output tokens, with cached reads at $0.015.
Was Space Bunny really MiniMax?
Yes. MiniMax said so on 19 October, and OpenRouter’s Space Bunny Alpha page now says the stealth model was developed and operated by MiniMax and revealed to be M3.1 Flash.
When will the M3.1 Flash weights be released?
MiniMax’s X post says 23 October 2026. The blog only says “later,” and MiniMax has not named a license yet, so check the Hugging Face page before building on it.
Is M3.1 Flash better than DeepSeek V4.1 Flash for coding?
Not on MiniMax’s own numbers. DeepSeek V4.1 Flash 1017 scores 77.0 on DeepSWE v1.1 against 73.7 for M3.1 Flash, and DeepSeek’s off-peak prices match MiniMax’s flat rate.
What happens to MiniMax M2.5 and M2.7?
They retire on 18 November 2026. API users must migrate before then, while M Plan and Token Plan requests are already routed to M3.1 Flash.