The anonymous “Space Bunny” model that developers have been running for free since late September finally has a name and a price tag. MiniMax said Monday morning that Space Bunny was an early version of MiniMax M3.1 Flash, and the MiniMax M3.1 Flash price on its API is a flat $0.15 per million input tokens, $0.60 per million output tokens and $0.015 for cached input.
Those headline numbers match DeepSeek’s V4.1 Flash 1017, which shipped two days earlier, and both models target the same work: long-context coding and agents that call tools in a loop. MiniMax’s launch charts make the case that its model is the better deal. Read the footnotes, though, and one of those charts bills DeepSeek at its peak-hour rate, which is double what most US developers pay.
MiniMax M3.1 Flash price: half of M3 per token, not a third
MiniMax’s pricing page, checked October 19, lists one flat rate for M3.1 Flash across its full context window. The bigger MiniMax M3, released in June, costs twice as much per token for prompts up to 512,000 tokens and four times as much above that.
| MiniMax API, per 1M tokens (Oct 19) | Input | Output | Cached input |
|---|---|---|---|
| M3.1 Flash (flat) | $0.15 | $0.60 | $0.015 |
| M3, prompts up to 512K | $0.30 | $1.20 | $0.06 |
| M3, prompts over 512K | $0.60 | $2.40 | $0.12 |
MiniMax’s launch post on X says M3.1 Flash runs “at 1/3 the cost of M3,” and its launch blog says overall usage cost on agentic tasks is 70% lower. Neither is the token price, which is 50% lower for most prompts. The rest of the claimed savings has to come from the model finishing tasks with fewer tokens and turns. That’s MiniMax’s measurement, on its own task mix.
A 1M-token context, video input, and thinking you can’t switch off
MiniMax’s developer docs list a 1,000,000-token context window and text, image and video input. The model ID is MiniMax-M3.1-Flash, and it works through both Anthropic-style and OpenAI-style endpoints. Reasoning depth is set with an effort parameter (low, medium, high, xhigh or max), and the default is max. OpenRouter’s model listing marks reasoning as mandatory and caps output at 524,288 tokens.
That default matters for the bill. MiniMax’s own Automation-Bench chart shows cost per task rising at every step from low to max effort, and its headline scores are max-effort runs. MiniMax says output can reach more than 200 tokens per second, with a median above 170. It hasn’t published a parameter count. For reference, M3’s Hugging Face card lists about 428 billion parameters with about 23 billion active, and MiniMax says Flash gets its speed from “more aggressive sparsity.”
Space Bunny was a public test drive
MiniMax says Space Bunny processed more than 200 trillion tokens while it ran anonymously. According to CellCog’s running record of the model, OpenRouter listed it as a free “stealth” model on September 23, and tokenizer tests by independent developers pointed to MiniMax within days. MiniMax confirmed it with a teaser video of a carrot-shaped rocket at 5:21 AM ET on Monday, followed by the full launch post at 7:56 AM ET.
If you formed an opinion during the free period, it was of a preview. Ryan Lee, MiniMax’s head of developer relations, replied on X that the final version is better at coding than Space Bunny was.
MiniMax’s own table puts DeepSeek ahead on coding
MiniMax published a 15-row comparison. All the figures below come from that table, run or collected by MiniMax. It doesn’t say who ran the rival scores, and its US rivals are Claude Opus 5 and GPT-5.6 Sol, not Anthropic’s or OpenAI’s newest models.
| MiniMax’s table (Oct 19) | M3.1 Flash | DeepSeek V4.1 Flash 1017 | GPT-5.6 Sol | Opus 5 |
|---|---|---|---|---|
| Automation-Bench v1.0.6 | 62.2 | 60.5 | 45.8 | 50.3 |
| DeepSWE v1.1 | 73.7 | 77.0 | 73.0 | 74.0 |
| Terminal Bench 4.0 | 38.1 | 42.3 | 39.9 | 51.8 |
| OSWorldV2 | 61.7 | not listed | 65.7 | 70.6 |
| HLE (no tools) | 43.4 | 42.6 | 49.5 | 55.6 |
On MiniMax’s own numbers, DeepSeek’s two-day-old model beats M3.1 Flash on the DeepSWE coding benchmark and on Terminal Bench. M3.1 Flash wins on Automation-Bench, which tests multi-step business workflows across SaaS tools. MiniMax’s separate Agents’ Last Exam chart also shows Claude Opus 5.5 and GPT-6 Astra scoring higher than M3.1 Flash’s 29.6%, at estimated costs above $1,000 for the 152-task run versus $111 for MiniMax.
The footnote on MiniMax’s cost chart
MiniMax’s Automation-Bench chart plots score against average cost per task. M3.1 Flash sits far to the left, cheaper than everything else. The footnote says DeepSeek V4.1 Flash 1017 was priced at peak rates: $0.30 input, $1.20 output and $0.006 cached.
DeepSeek’s pricing page sets peak hours at 01:00 to 04:00 and 06:00 to 10:00 UTC on weekdays. On US Eastern time, that’s overnight and early morning, so a normal American workday is off-peak, at $0.15 input and $0.60 output, the same as MiniMax. DeepSeek’s off-peak cached input is $0.003, a fifth of MiniMax’s $0.015, which matters for agents that resend the same long context on every turn. Halve DeepSeek’s prices and its max-effort point on that chart lands at roughly the same cost per task as M3.1 Flash’s, by our reading of the chart. MiniMax still shows the higher score, though its own materials disagree slightly on what it is: 62.2 in the table, 62.8% on the chart.
Weights are due Thursday, if the X post is right
MiniMax’s X post says the model weights “will arrive on Oct 23.” The blog says only that they will be “released later,” and the docs already describe M3.1 Flash as open-weight. MiniMax hasn’t said what license the weights will use. M3 shipped under a custom MiniMax community license, not MIT or Apache 2.0, and DeepSeek’s Flash weights are already out under MIT.
Nobody outside MiniMax has published a full benchmark run of the final model yet. OpenRouter lists M3.1 Flash, but its endpoint list was empty when we checked Monday morning, so for now the ways in are MiniMax’s API, its M Plan subscription and its MiniMax Code app. A larger “Pro” model will “arrive later,” MiniMax says, with no date. The next check is Thursday: whether the weights land on Hugging Face, and under what license. We’ll update this story and our LLM News feed when they do.
Frequently asked questions
How much does MiniMax M3.1 Flash cost?
On MiniMax’s API pricing page, checked October 19, 2026, M3.1 Flash costs $0.15 per million input tokens, $0.60 per million output tokens and $0.015 per million cached input tokens. It’s a flat rate across the whole 1-million-token context, with no long-prompt surcharge.
What was Space Bunny?
Space Bunny was an anonymous model that appeared free on OpenRouter in late September. MiniMax confirmed on October 19 that it was an early version of M3.1 Flash, and its head of developer relations says the final release codes better than Space Bunny did.
Is MiniMax M3.1 Flash open source?
Not yet. MiniMax’s X post says the weights will arrive on October 23, while its blog only says “later.” The license hasn’t been announced; MiniMax’s previous model, M3, uses a custom MiniMax community license.
Is MiniMax M3.1 Flash cheaper than DeepSeek V4.1 Flash 1017?
Not on list price during US business hours. Both charge $0.15 input and $0.60 output per million tokens, and DeepSeek’s off-peak cached input is cheaper at $0.003. DeepSeek doubles its prices during its peak hours (overnight and early morning on US Eastern time), while MiniMax charges one flat rate around the clock.
What is MiniMax M3.1 Flash’s context window?
One million tokens, according to MiniMax’s docs, with text, image and video input. OpenRouter lists a maximum output of 524,288 tokens. Thinking can’t be turned off, and the default effort level is max.
Sources
All accessed October 19, 2026, between about 8:45 and 9:00 AM ET.
- MiniMax, “MiniMax M3.1 Flash: Redefining Agent Performance and Cost” (Oct 19, 2026; Space Bunny confirmation, 200T+ tokens claim, benchmark claims, 70% lower usage cost claim, 200+ TPS claim, flat pricing, weights “later,” Pro “later”): minimax.io/blog/minimax-m31f
- MiniMax API docs, Pricing Overview (M3.1 Flash $0.15 / $0.60 / $0.015; M3 standard and Priority tiers): platform.minimax.io/docs/guides/pricing
- MiniMax API docs, Text Generation (model ID, 1,000,000-token context, text/image/video input, effort levels and max default, Anthropic- and OpenAI-compatible endpoints): platform.minimax.io/docs/guides/text-generation
- @MiniMax_AI on X, launch post (Oct 19, 2026, 7:56 AM ET; “1/3 the cost of M3,” weights “will arrive on Oct 23”), benchmark table and cost charts: x.com/MiniMax_AI/status/2112150867868541221 and x.com/MiniMax_AI/status/2112152630617337952
- @MiniMax_AI on X, “Welcome back. Space Bunny.” teaser video (Oct 19, 2026, 5:21 AM ET): x.com/MiniMax_AI/status/2112111925975032241
- Ryan Lee (MiniMax head of developer relations) on X, reply on the final version vs Space Bunny (Oct 19, 2026): x.com/RyanLeeMiniMax/status/2112132626211598467
- OpenRouter model listing, MiniMax M3.1 Flash (1,000,000 context, 524,288 max output, mandatory reasoning; endpoints checked Oct 19): openrouter.ai/minimax/minimax-m3.1-flash
- MiniMax-M3 model card on Hugging Face (about 428B total / 23B active parameters; MiniMax community license): huggingface.co/MiniMaxAI/MiniMax-M3
- DeepSeek API docs, Models & Pricing (V4.1 Flash off-peak and peak rates, cache-hit price, peak hours in UTC): api-docs.deepseek.com/quick_start/pricing
- CellCog, “Space Bunny Alpha Is MiniMax M3.1-Flash: The Record” (OpenRouter stealth listing Sep 23, tokenizer tests): cellcog.ai/blog/what-is-space-bunny-alpha
Related: DeepSeek V4.1 Flash 1017 price and benchmarks · Reflection Beam open weights · More LLM news
Last updated: October 19, 2026. To report an error, see our corrections page. Articles are drafted with AI assistance and reviewed and edited by an editor; see our editorial policy.
