On the evening of October 19, MiniMax officially released M3.1 Flash, the latest generation of its Flash model. The company's blog post confirmed something the industry had speculated about for months: Space Bunny, the anonymous model that had processed more than 200 trillion tokens, was an early version of M3.1 Flash.

MiniMax describes the capability jump as "generational" — matching leading models on complex agent tasks, computer use, software development and cybersecurity, while running inference nearly three times faster and cutting total usage cost by 70% versus the previous generation. Domestic API pricing came down in step: 1 yuan per million input tokens, 4 yuan per million output tokens, and 0.1 yuan per million cached tokens read.

An Anonymous Test Spanning 200 Trillion Tokens

How Space Bunny operated deserves its own note. The model served the open market under an anonymous identity, and only after it had processed more than 200 trillion tokens did MiniMax confirm it was an early version of M3.1 Flash.

That amounted to an unusually large public blind test. A call volume of 200 trillion tokens means the model was stress-tested at scale in real production environments, not just tuned to score well on closed benchmark sets. For a model company, the cost of this approach is giving up brand exposure during the anonymous period; the payoff is usage data no third party can fake.

MiniMax positions the new model as reaching "the level of leading models" across four task categories: complex agent work (cowork), computer use, software development (coding) and cybersecurity.

Benchmark Results Across Four Areas

According to MiniMax's official blog, M3.1 Flash posted specific scores on several benchmarks:

  • Complex tasks and automation: top-tier results on AutomationBench-AA; in real enterprise workflows spanning SaaS systems, it can fully resolve end-to-end problems across sales, marketing, operations, customer service, finance and HR; a score of 76.2 on BankerToolBench, with strengths in multi-step planning, cross-document consistency, financial compliance and chained tool calls.
  • Programming and software engineering: 73.7% on DeepSWE v1.1, which MiniMax calls "on par with top models," highlighting performance on long-horizon tasks.
  • Computer use: 61.7 on OSWorld V2 and 29.6 on Agents' Last Exam (ALE), which MiniMax says ranks first among Chinese models on both benchmarks.
  • Cybersecurity: 82% on CyberGym, with MiniMax claiming leading performance in code auditing, vulnerability patching, penetration testing, malware analysis and reverse engineering.

Worth noting: all of these benchmarks were selected and reported by MiniMax itself, with no independent third-party reproduction. The "first place" claims on OSWorld V2 and ALE refer only to rankings among Chinese models, not globally.

Price and Speed: The Cost and Payoff of Aggressive Sparsity

M3.1 Flash's efficiency gains come from architectural changes. MiniMax says it "further optimized inference efficiency through more aggressive sparsity," which Chinese media rendered as a "more aggressive sparse attention architecture." On the user side, the company reports output speeds of 200+ TPS in everyday scenarios, with a median (P50) holding steady above 170 TPS.

Two sets of pricing figures need to be kept apart. MiniMax's dollar pricing is $0.15 per million input tokens, $0.60 per million output tokens and $0.015 per million cached input tokens. Domestic API pricing has come down to 1 yuan per million input tokens, 4 yuan per million output tokens and 0.1 yuan per million cached tokens read.

According to Shanghai Securities News, compared with the equivalent M3 pricing (inputs of 512K tokens or less), both input and output unit prices fell by more than 50%, while the cached-read price dropped by more than 70%. The "70%" in MiniMax's English blog post, however, refers to a 70% reduction in total usage cost versus the previous generation — a figure tied to the overall expense of agentic workloads, not to unit price cuts. Both numbers are accurate, but they measure different things.

The model is currently available through three channels: the MiniMax API, M Plan and MiniMax Code. MiniMax states that "model weights will be released later," meaning it is not open source for now. A Pro version aimed at a higher intelligence ceiling is teased for release "soon," with no timeline given.

Why Chinese Flash Models Are All Chasing Agent Workloads

Set M3.1 Flash alongside a few recent moves and the direction is consistent: DeepSeek V4.1 Flash was updated to V4.1-Flash-1017 on October 17, leaning on self-feedback reinforcement learning to improve its ability to build complete projects from scratch in areas like game development and 3D modeling; StepFun open-sourced the weights for Step 5 Preview on October 15, likewise pitching agentic tasks as the core selling point.

All three point to the same conclusion: the main battleground on the inference side has shifted from "single-turn answer quality" to "multi-step task completion rate." AutomationBench-AA, DeepSWE, OSWorld and ALE are not measuring the quality of a response, but whether a model can hold to its objective after dozens of operations.

The simultaneous cooling and restarting of the price war is also happening at this moment. M3.1 Flash pushed unit prices down to 1 yuan for input and 4 yuan for output, while StepFun, three days earlier, doubled the cache-hit input price for Step 5 Preview from 0.35 yuan to 0.70 yuan. One company cutting prices by more than 50% and another raising them by 100% in the same month suggests this round of pricing divergence is no longer a straightforward price war, but a set of choices different models are making about their own cost structures and target customers.