MiniMax Unmasks Space Bunny as M3.1 Flash

MiniMax unmasked Space Bunny on October 19, 2026, naming the anonymous model that had already processed more than 200 trillion tokens as M3.1 Flash. The company put the model on its API, M Plan, and MiniMax Code, and said it is reducing usage costs by 70% on agentic jobs while making online inference nearly three times faster.

That is a price tag on traffic that already existed, not a cold start. The wager is that a cheap worker for coding, computer use, and office agents will pull more volume than another flagship chat model.

The Anonymous Run That Burned 200 Trillion Tokens

OpenRouter listed Space Bunny Alpha at 14:48 UTC on September 23, 2026, as an anonymous large model with a 1 million token context, multimodal input, and a provider field of Stealth. Tokenizer probes the same week matched MiniMax-family token counts on 50 of 50 strings, which is why the guess landed on MiniMax days before the company said a word.

Four days after that listing, MiniMax put MiniMax-M3.1-Flash-Preview inside MiniMax Code and on its Token Plan, still without a public per-token price or a score table. The stealth route was pulled on October 5, 2026. On October 19 the lab closed the loop.

Space Bunny, which processed more than 200 trillion tokens while operating anonymously, was an early version of MiniMax M3.1 Flash.

MiniMax, M3.1 Flash launch post, October 19, 2026

Blind usage is the cleaner demand signal here than a vendor scorecard. A model does not chew through 200 trillion tokens in the dark unless coding agents and routers are already feeding it real work.

THE SPACE BUNNY CALENDAR

1. September 23, 2026: Space Bunny Alpha appears on OpenRouter as an unnamed Stealth model.
2. September 27, 2026: MiniMax-M3.1-Flash-Preview opens in MiniMax Code and on the Token Plan.
3. October 5, 2026: The anonymous OpenRouter listing comes down.
4. October 19, 2026: MiniMax names Space Bunny as early M3.1 Flash and opens the paid API.

The official account posted the same unmasking in public, with weights promised three days later.

MiniMax M3.1 Flash: Space Performance + Bunny Speed

🚀🐰 Yes, Space Bunny – which has topped global leaderboards with 200T+ tokens – was an earlier version of M3.1 Flash.

💰Redefining the performance economics of agentic AI. M3.1 Flash achieves frontier-level intelligence… pic.twitter.com/j0tCfNYYgo

— MiniMax (official) (@MiniMax_AI) October 19, 2026

Flash Now Costs $0.15 In and $0.60 Out

Pay-as-you-go customers now see flat rates of $0.15 and $0.60 per million tokens, with cached input at $0.015. The Chinese list is 1 yuan in, 4 yuan out, and 0.1 yuan for cache reads, per million tokens. MiniMax-M3 stays on the menu at $0.30, $1.20, and $0.06 for prompts up to 512,000 input tokens, and at $0.60, $2.40, and $0.12 above that line.

M3.1 FLASH API RATES

Those sticker cuts are 50% on input, 50% on output, and 75% on cache reads, which is not the same number as the 70% agentic-task claim. Agent jobs reread long prompts. OpenCode’s usage tables on this model show a 96% cache ratio, so the cheap cache line does more work than the headline input rate. Public price lists put DeepSeek’s V4.1 Flash at the same $0.15 and $0.60 off-peak, with a peak window that doubles those rates in Beijing business hours; MiniMax does not split the day.

Speed is the other half of the bet. MiniMax says everyday output can hit 200+ tokens per second, with a median (P50) of 170+ tokens per second, after more aggressive sparsity at inference. Ryan Lee, MiniMax’s head of developer relations, wrote “Enjoy the speed!” under the launch thread. Teams that leave effort at the default will still pay for a long thinking stream before any of that speed shows up in the answer.

A 3.3-Point DeepSWE Gap on MiniMax’s Chart

MiniMax’s launch table puts M3.1 Flash next to frontier names on coding, office agents, computer use, and visual reasoning. The company also left a rival on that table. DeepSeek V4.1 Flash-1017 is listed at 77.0% on DeepSWE v1.1 beside Flash’s 73.7%, a 3.3-point gap MiniMax did not hide.

MINIMAX’S OWN SCORECARD

Flash ranks first among Chinese models on OSWorld V2 and Agents’ Last Exam, MiniMax said. BankerToolBench is the office-finance stress test: chained tools, document consistency, and compliance. Automation-Bench is the cowork row, the one MiniMax is selling to sales, marketing, operations, customer service, finance, and HR teams that want an agent to finish a SaaS workflow rather than draft an email about it. None of these figures has a public third-party rerun attached to the launch, so they sit as the company’s own board.

What Happens When Thinking Cannot Be Turned Off?

M3 let developers switch reasoning off. M3.1 Flash will not. Thinking is on for every request, and the only control is how hard it thinks. Send thinking.type=disabled or effort: "none" and the API returns HTTP 400 with the line that the model requires adaptive thinking. The docs tell you to lower effort if you want less latency, not to hunt for an off switch.

THE FIVE EFFORT LEVELS

• low: Shortest thinking stream, least extra output, weakest on hard agent jobs.
• medium: A middle setting for routine tools and smaller patches.
• high: Deeper planning for multi-file coding and longer tool chains.
• xhigh: Heavier reasoning before the visible answer starts.
• max: The default if you omit the field, and the most expensive in thinking tokens.

On the Anthropic-compatible endpoint the field is output_config.effort. On the OpenAI-compatible endpoint it is reasoning_effort. Omit it and you get max. That default is easy to miss, and thinking tokens count against max_tokens, so a low cap can come back as an empty answer after the model has spent the budget on a hidden chain of thought. For agent loops that already reread a repo, max on every call is how a “cheap” Flash bill climbs.

Retiring M2.5 Forces Traffic Onto Flash

The launch is also a migration. MiniMax closed M2.5 and M2.7 to new pay-as-you-go users on October 19, 2026, and will retire those families on November 18, 2026. Existing metered users can keep calling them until that date. Under M Plan and Token Plan, requests to M2.5 and M2.7 already route to M3.1 Flash, with old model names still accepted so clients do not have to be rewritten overnight.

That is a hard shove toward the new default. Anyone who liked M2.7-highspeed at about 100 tokens per second, or who had turned M3 thinking off to save money, now sits on a model that always thinks and bills Flash rates. Plan quotas are unchanged. The behavior is not. MiniMax-M3 remains available for teams that want the older control surface and will pay $0.30 and $1.20 to keep it.

Computer Use, Cyber Work, and a 1M Window

The product MiniMax is selling is not a chatbot with a discount. It is a worker that is supposed to click, type, and call tools until a job is done. The OSWorld 2.0 computer-use benchmark is built around 108 long desktop workflows that take a skilled human a median of about 1.6 hours, which is the class of task Flash’s 61.7 score is meant to speak to. Agents’ Last Exam at 29.6 is the paid-work version of the same idea.

Cybersecurity is the odd extra on a Flash card. MiniMax reports 82% on CyberGym for vulnerability discovery and validation, and says its own security reviewers used the model on code audits, patching, penetration tests, malware analysis, and reverse engineering. Those are in-house checks, not a public board. They still tell you where the company wants cheap tokens to go: long, tool-heavy jobs that used to be too spendy to leave running.

Context is 1,000,000 tokens, shared with M3, for whole repositories and multi-step agent sessions. Input is text, images, and video. The open MiniMax-M3 weights on Hugging Face describe a mixture-of-experts model with about 428 billion parameters and about 23 billion active, and Flash’s preview developer guide listed the same 428 billion total. MiniMax credits the extra speed to sparser inference, not a smaller net.

Open Weights Arrive on October 23

M3.1 Flash is live on the MiniMax API at platform.minimax.io, on M Plan, and in the MiniMax Code desktop app. Model weights arrive on October 23, 2026. Until then the only way to run it is as a hosted service. A Pro model, which MiniMax says will push the intelligence ceiling, has no date and no price.

The hosted bet is already on the table at $0.15 and $0.60, with a cache line that makes long agent sessions look cheaper than the input rate implies. The open-weight bet is the next three days. If those checkpoints land, the same worker that ran anonymously through 200 trillion tokens can sit on someone else’s GPU at Flash economics. If they slip, the API price is still the product, and DeepSeek is already standing on that same sticker.

Frequently Asked Questions

What Was Space Bunny?

Space Bunny Alpha was the unnamed OpenRouter and OpenCode listing that appeared on September 23, 2026, billed as a free preview from a Stealth provider, with a 1 million token window and text, image, and video input. MiniMax said on October 19, 2026 that this anonymous model was an early version of M3.1 Flash and that it processed more than 200 trillion tokens before the name went on.

Can You Turn Off Thinking on MiniMax M3.1 Flash?

No. The text-generation docs state that thinking is on by default and cannot be disabled, and that sending a disable flag returns HTTP 400. The supported fix is to set effort to low or medium, which shortens the hidden reasoning stream and cuts latency without turning reasoning off.

What Effort Levels Does MiniMax M3.1 Flash Support?

Five: low, medium, high, xhigh, and max. If the field is omitted, the default is max on both the Anthropic-compatible and OpenAI-compatible APIs. Higher levels emit more thinking tokens and raise latency, which is why a coding agent that forgets to set effort can burn a Flash budget as if it were still on max.

Which APIs Accept MiniMax M3.1 Flash?

MiniMax documents an Anthropic-compatible base at https://api.minimax.io/anthropic, which it calls the default path because it keeps thinking blocks, and an OpenAI-compatible base at https://api.minimax.io/v1 for teams that do not want to change SDKs. The model id is MiniMax-M3.1-Flash on both, and the same id is what M Plan and MiniMax Code send when they route old M2 names forward.