On October 19, 2026, MiniMax officially released M3.1 Flash and revealed that the model previously running anonymously as "Space Bunny" was its early version. According to MiniMax's official blog, Space Bunny processed over 200 trillion tokens in anonymous status; on OpenRouter alone, it processed 56.7 trillion tokens and held the top spot in platform usage for 10 consecutive days. Inference speed increased nearly 3x, comprehensive usage cost dropped 70%, and OSWorld V2 score of 61.7 ranked first among Chinese models—these are the three core numbers presented at M3.1 Flash's launch.
MiniMax chose the path of running anonymously first and officially debuting later. OpenRouter, as a multi-model intermediary platform, aggregates a large number of real developer calls, covering a wide range of scenarios with dense token consumption, and can serve as a natural test bed close to production environments. Running anonymously as "Space Bunny" avoids brand interference and collects real usage data. Compared with the early version, first-token latency dropped from 1.95 seconds to 1.22 seconds (38% improvement), and throughput increased from 70 tokens per second to 100 tokens per second (45% improvement). At official launch, MiniMax claimed peak output speed exceeding 200 TPS, with P50 median speed stable above 170 TPS. Artificial Analysis measured 116.2 TPS, ranking 78th among 151 tested models.
Pricing Logic: Where It's Cheap and How Much Cheaper
M3.1 Flash is priced at $0.15 per million input tokens, $0.60 per million output tokens, and $0.015 per million cached read tokens. The current mainstream model market median price is $0.24 input and $0.87 output. MiniMax's emphasized 70% cost reduction targets "comprehensive usage cost for agent-type tasks": agent scenarios rely heavily on cached input, and the total number of tokens required to complete the same goal decreases, with the two factors combining to create the cost difference.
Domestic pricing is 1 yuan per million input tokens and 4 yuan per million output tokens. The China region and USD region have different pricing systems, and developers should calculate based on the actual API region quotation.
Capability Structure: Where It's Strong and What the Numbers Say
MiniMax provides six benchmark results. OSWorld V2 score of 61.7 ranks first among Chinese models. BankerToolBench score of 76.2. DeepSWE v1.1 score of 73.7%. CyberGym score of 82%. Agents' Last Exam (ALE) score of 29.6. AutomationBench-AA score of 72.5%.
Artificial Analysis ranks M3.1 Flash 4th among 173 tested models, with an intelligence index score of 43 (median among tested models is 13). M3.1 Flash's model weights are not currently available for download; MiniMax promises to open them "later" but has not given a specific date. All calls must go through MiniMax API, MPlan, or MiniMax Code cloud access.
Chinese Models' "Low-Price Benchmarking" Has Become Systematic
Over the past two years, DeepSeek released high-performance models at prices far below OpenAI, triggering a global repricing of AI. M3.1 Flash's high scores are concentrated in agent tasks, tool calling, and computer operation tracks rather than comprehensive general language capability. MiniMax positions M3.1 Flash as a Flash-series product balancing "efficiency and capability" and teases that a Pro version "pursuing the upper limit of intelligence" will be released later.
Practical Advice for Developers and Enterprise Selection
M3.1 Flash has reference value for three types of scenarios: developers of multi-step agent applications can test the combination of OSWorld V2 and BankerToolBench scores with cached pricing; engineering teams with programming assistance or code review needs can note DeepSWE v1.1's 73.7% score; scenarios requiring private deployment need to wait for weight availability.
Three points need verification before use: there is a gap between the officially claimed 200+ TPS and Artificial Analysis's measured 116.2 TPS, so latency-sensitive applications should complete their own stress testing; the 70% cost reduction is based on comprehensive calculation for agent scenarios, and actual cost advantages in simple batch processing or non-cached scenarios will be lower than this figure; China region and USD region pricing differ, and calculations must be done separately by calling region.
From a product cycle perspective, the Pro version teased by MiniMax is a key node for testing the sustainability of the technical路线. If the Pro version can compete head-on with top models on general reasoning, then the Flash/Pro dual-line layout has strategic completeness; if the Pro version is delayed long-term or only leads in vertical categories, then M3.1 Flash's release is more of a pricing-driven market move. Weight availability timing is another signal: the faster they open, the more MiniMax's judgment of its own moat leans toward ecosystem positioning; if it maintains cloud closed-source long-term, its business model still depends on API service barriers.
© 2026 Winzheng.com 赢政天下 | 转载请注明来源并附原文链接