MiniMax Reveals M3.1 Flash Was Secret 'Space Bunny' That Processed 200 Trillion Tokens
MiniMax rolls out M3.1 Flash, a 1M-context agentic coding model priced at $0.15/$0.60 per million tokens, with weights promised soon.
- MiniMax launched M3.1 Flash, confirming it was the stealth Space Bunny model on OpenRouter.
- Pricing: $0.15 per million input tokens, $0.60 output, $0.015 cached.
- 1M-token context, multimodal input (text, image, video), always-on reasoning with effort levels.
- Company reports 73.7% DeepSWE v1.1, 61.7 OSWorld V2, 82% CyberGym.
- Claims 70% lower agent task cost and nearly 3x faster inference versus M3.
- Available via OpenRouter, MiniMax API, M Plan, and MiniMax Code; weights promised October 23.
MiniMax has released M3.1 Flash, a lower-cost sibling to M3 designed for coding, research, computer use, and multi-step agent workflows. The company also identified it as Space Bunny, the anonymous model that accumulated heavy traffic during a stealth deployment on OpenRouter.
MiniMax says Space Bunny processed more than 200 trillion tokens during testing. OpenRouter reports that 56.7 trillion of those tokens passed through its platform, where the model ranked first by usage for 10 consecutive days. The traffic provides evidence of sustained demand and operational scale, though it does not independently verify MiniMax’s performance claims.
Pricing favors long agent runs
| Usage | Price per million tokens |
|---|---|
| Input | $0.15 |
| Cached input | $0.015 |
| Output | $0.60 |
A request using one million uncached input tokens and one million output tokens would cost $0.75 before tool or infrastructure charges. MiniMax’s full M3 model costs $0.30 per million input tokens and $1.20 per million output tokens, making M3.1 Flash half the price at the published token rates.
MiniMax also claims that M3.1 Flash cuts task-level costs by 70% compared with its predecessor and runs online inference nearly three times faster. Task-level cost includes the tokens consumed across tool calls, retries, planning steps, and intermediate outputs. That measure is especially relevant for agents, where the number of loops can affect the final bill as much as the listed token price.
A large context window and familiar APIs
- Context: Up to one million tokens, allowing a request to include large repositories, document collections, or extended interaction histories.
- Output: Up to 524,288 tokens per response.
- Modalities: Native image and video input with text output.
- Reasoning: Always enabled, with adjustable effort levels that let developers trade latency and cost for additional computation.
- API access: OpenAI-compatible and Anthropic-compatible endpoints, plus M Plan and MiniMax Code.
- Hosted access: A free OpenRouter route is available during the launch period.
- Weights: MiniMax’s announcement lists October 23 as the release date for model weights. The company has not published a license.
API compatibility should reduce migration work for applications already using OpenAI or Anthropic client libraries. Teams should still test streaming, tool schemas, error handling, rate limits, and structured output because compatible endpoints do not guarantee identical behavior.
Benchmarks target agent work
MiniMax reports its strongest results in autonomous software engineering, desktop control, and cybersecurity. Each benchmark tests a multi-step workflow in which the model must inspect an environment, choose actions, and respond to the results.
| Benchmark | Reported score | What it measures |
|---|---|---|
| DeepSWE v1.1 | 73.7% | Repository-level software work, including reading code, implementing changes, fixing bugs, and running tests |
| OSWorld V2 | 61.7 | Desktop operation through screenshots, clicks, keyboard input, and application state |
| CyberGym | 82% | Security tasks such as vulnerability discovery and exploitation |
These results come from MiniMax and have not yet been independently reproduced. The company previously reported 80.5% for M3 on SWE-bench Verified, but that benchmark differs from DeepSWE v1.1, so the two scores do not support a direct quality comparison between the models.
Three limits to verify
- Task-cost methodology: The claimed 70% reduction depends on complete agent runs. Results may vary with prompts, tool design, retry policies, and reasoning effort.
- Independent evaluation: Public testing is still needed to confirm benchmark scores, latency, reliability, and long-context performance under production workloads.
- Weight licensing: M3.1 Flash remains proprietary at launch. A weight release would enable self-hosting only if the eventual license permits the intended commercial, research, or modification rights.
The Space Bunny deployment shows that the model has already handled substantial hosted traffic. It leaves open questions about request mix, failure rates, routing conditions, and performance on tasks outside MiniMax’s selected benchmarks.
Where M3.1 Flash fits
Teams building coding agents, browser and desktop agents, or long-running research systems have the clearest reason to evaluate M3.1 Flash. Its one-million-token context window can hold large working sets, while lower token prices and adjustable reasoning effort offer direct controls over operating cost.
A useful bake-off should measure end-to-end task success, wall-clock time, total input and output tokens, tool calls, retries, and cost per completed task. Developers should also test long-context retrieval, image and video handling, structured outputs, and compatibility with existing tool definitions.
For chat and short single-turn generation, benchmark quality and latency against the current stack because MiniMax’s largest savings claim applies to multi-step agents. For workloads dominated by repeated reasoning and tool use, M3.1 Flash combines aggressive pricing with evidence that it can sustain significant production traffic.
Comments
