Space Bunny Was MiniMax M3.1 Flash, and It Costs DeepSeek's Off-Peak Price All Day. You Cannot Turn Its Thinking Off.

For ten days an anonymous model called Space Bunny was the most used model on OpenRouter, and for twelve days it was the most used model on OpenCode. This morning, October 19, 2026, at 6:56 AM CT, MiniMax posted the reveal: Space Bunny "was an earlier version of M3.1 Flash," and M3.1 Flash is now on MiniMax's pay-as-you-go API at a flat $0.15 in and $0.60 out per million tokens.

That price is the exact off-peak rate of DeepSeek V4.1 Flash, the model that got rescored to 45 on Artificial Analysis's index two days ago, except MiniMax has no peak window. Artificial Analysis scored M3.1 Flash today at 42, fourth of 173 in its price class, at $0.22 per index task. Three points under DeepSeek at the same money, all day, every day.

Then the fine print. Thinking cannot be turned off and defaults to max, so every call pays for reasoning tokens unless you say otherwise. MiniMax's blog says median speeds of "170+ TPS"; AA measured 120. And anyone on an M Plan calling the old M2.5 or M2.7 model names is now served M3.1 Flash, the same same-name pattern DeepSeek ran two days ago. We read the launch post, the pricing page, the text-generation guide, the model-retirement FAQ, AA's page, and the partner posts. Here is what held up.

What You Are Calling

Half of M3 per token, not 70% off. MiniMax's own M3 standard tier is $0.30 in and $1.20 out up to 512K tokens, so M3.1 Flash is exactly half per token and a quarter on cache reads. The blog's "70% lower" usage cost for agentic tasks is MiniMax's claim with no workload published; the per-token arithmetic is 50%.

The thinking is the bill. MiniMax's docs: "Thinking cannot be turned off." Sending thinking: {"type": "disabled"} or effort: "none" returns a 400 with the message that the model "requires adaptive thinking." Omit effort and you get max. AA's numbers show what max costs: 180M output tokens to run its index against a class median of 99M, which AA labels "very verbose."

The Benchmarks, Sorted by Who Reported Them

House rules: independent first, with the source; MiniMax's numbers labeled as MiniMax's.

Artificial Analysis, Intelligence Index v4.3.2, October 19. 42, fourth of 173 in its price class. Output speed 120.1 tokens per second on MiniMax's API, 70th of 151 in that class, against a median of 116. Cost per index task $0.22. Output tokens to run the index 180M. AA ranks it in a price class, not the open-weight class, and lists no license or parameter count, because as of today there are no weights.

Now the vendor table. These are MiniMax's self-reported numbers from the launch post.

Three honest observations:

• The speed claim and the measurement disagree. MiniMax: output "can reach 200+ TPS" with "median (P50) speeds consistently reaching 170+ TPS." AA on the same API: 120.1. Launch-day load may explain some of it. Print both.
• The launch-week usage numbers are inflated by free access. OpenRouter's free window runs 72 hours from about 7 AM CT today; OpenCode's runs three days; Vercel's Gateway lists the model at $0. The 200 trillion tokens MiniMax says Space Bunny processed while anonymous were not all paid tokens.
• Secondary sites print chart numbers we could not read. AICoder cites Automation-Bench 62.2%, BabyVision 91.0%, and Terminal-Bench 4.0 38.1%. None of those appear in the blog text; they are likely read off chart images. We do not print them.

Same Name, New Model, Again

MiniMax's model FAQ says the M2.5 and M2.7 families closed to new users today and retire on November 18, 2026, and then the sentence that matters: "Starting October 19, 2026, all requests to M2.5 and M2.7 models under these plans will automatically be routed to M3.1 Flash." That covers the M Plan and Token Plan. The old model names keep working; they just return a different model. Pay-as-you-go users keep the old models until November 18, after which those requests stop. The same page's retired-models table already lists M2.7 and M2.5 as retired, which contradicts its own dates.

Two Chinese labs, one week, one pattern: DeepSeek moved deepseek-flash to a new build on October 17; MiniMax is moving two model families onto a third under their old names today. If you are on an M Plan with M2.5 or M2.7 in your config, assume you are on M3.1 Flash now and re-run your evals.

Verified vs Unconfirmed: The Scorecard

Who Should Use It

Use M3.1 Flash for high-volume agentic work where a flat $0.15 and $0.60 with no peak window beats DeepSeek's two-tier clock, and set effort explicitly on every call; low or high for routine work, never the default. Compare it against DeepSeek Flash on your own tasks, because the per-token prices are identical off-peak and the AA scores are three points apart.

Skip it if you need thinking off, if your latency budget assumed MiniMax's 170 tokens per second rather than AA's 120, or if your M Plan config still says M2.5 or M2.7 and you have not re-tested since this morning.

The frontier is elsewhere: M3.1 Flash at 42 sits three under DeepSeek Flash 1017 at 45 and well under the closed flagships at the top of our Index. The story is the price class, where it is fourth of 173. Our model comparison has the current standings side by side.

Update, October 25, 2026

• October 22: The weights shipped. MiniMaxAI/MiniMax-M3.1-Flash is on Hugging Face with an NVIDIA NIM card dated October 22: about 428B total parameters and 23B active, five effort levels with thinking still required, and quantization notes for MXFP8 and NVFP4 experts. The license is the MiniMax Community License, not an open-source license: free for non-commercial use; commercial use must display "Built with MiniMax M3.1-Flash" and send a one-time notice to api@minimax.io; products above $20 million in yearly revenue need MiniMax's prior written authorization; military use is banned in the appendix. Some secondary sites print the attribution string as "Built with MiniMax M3"; the LICENSE file says M3.1-Flash.
• October 23: Artificial Analysis rescored the model from 42 to 43. The same day, vLLM published day-zero support on NVIDIA Blackwell, NVIDIA Rubin (with dedicated kernels), and AMD MI350-series, with prebuilt images (vllm/vllm-openai:minimax-m31-flash and the ROCm equivalent), a recommended serving shape that includes tensor parallelism across four GPUs, an NVFP4 KV cache, and DSpark speculative decoding; the PyPI wheel lands about a week after the upstream merge. The throughput and cost-per-token figures on that post are vLLM's own benchmarks. This is a multi-GPU datacenter model, not a laptop one.

Sources

• MiniMax: M3.1 Flash launch post
• MiniMax platform: pricing overview
• MiniMax platform: text generation guide
• MiniMax platform: model FAQ and retirements
• Artificial Analysis: MiniMax M3.1 Flash
• OpenRouter on X: Space Bunny was MiniMax M3.1 Flash
• OpenCode on X: M3.1 Flash free for three days
• Hugging Face: MiniMaxAI/MiniMax-M3.1-Flash
• Hugging Face: MiniMax Community License
• vLLM: day-zero support for MiniMax M3.1 Flash

*Not sure whether a flat-rate Flash beats a two-tier one for your traffic? Take the free 2-minute quiz and get matched. Then read the model it is priced against, and the same-name swap it is now copying: DeepSeek V4.1 Flash 1017.*