MiniMax has officially launched M3.1 Flash, revealing that its anonymous "Space Bunny" model — which quietly processed over 200 trillion tokens in the wild — was an early version of this release. The company positions M3.1 Flash as a generational leap in agentic tasks, coding, computer use, and cybersecurity, with online inference nearly three times faster and a 70% reduction in usage costs versus its predecessor. The benchmark numbers are specific and ambitious. M3.1 Flash scores 73.7% on DeepSWE v1.1 for complex software engineering tasks, 61.7 on OSWorld V2 for computer operation, 29.6 on Agents' Last Exam (ALE), 76.2 on BankerToolBench for multi-step enterprise workflows, and 82% on CyberGym for vulnerability discovery. MiniMax claims top-performer status on AutomationBench-AA and first place among Chinese models on OSWorld V2 and ALE. Output speeds reach 200+ tokens per second at peak, with median P50 speeds above 170 TPS. The pricing is the sharp edge. Flat rates of $0.15 per million input tokens, $0.60 per million output tokens, and $0.015 per million cached input tokens place M3.1 Flash at the aggressive low end of the inference market. MiniMax frames this as "redefining capability and cost for agentic tasks," and the math isn't wrong: if these benchmarks hold under real workloads, the cost-per-task for enterprise automation drops by an order of magnitude compared to frontier pricing from OpenAI or Anthropic. The strategic play is transparent. MiniMax is following the Chinese hyperscaler playbook: commoditize the inference layer, drive adoption through pricing, and lock developers into the ecosystem via MiniMax Code (a desktop agent), the M Plan subscription, and a forthcoming open-weight release. The 200-trillion-token stealth deployment as Space Bunny served as a live stress test at scale — a clever de-risking strategy that most Western labs haven't attempted so publicly. The gaps deserve scrutiny. All benchmarks cited are self-reported. There is no independent third-party audit, no mention of pre-registration, and several benchmarks (BankerToolBench, CyberGym, the internal security evaluation) are either proprietary or evaluated by MiniMax's own team. The cybersecurity claims — code auditing, penetration testing, malware analysis, reverse engineering — are assessed by "internal security experts," which is the weakest possible validation regime for a domain this sensitive. The model's "aggressive sparsity" for inference speed is mentioned without architectural detail. The broader market signal is clear: the floor for capable agentic models is collapsing. MiniMax joins DeepSeek in demonstrating that Chinese labs can match or approach Western frontier performance at dramatically lower price points. This compresses margins for API providers globally, benefits developers and enterprises building on these models, and raises the stakes for any company whose business model depends on inference pricing power. The announcement that a "Pro" model pushing intelligence boundaries will follow suggests MiniMax is building a two-tier product line — Flash for cost-sensitive volume, Pro for capability-intensive use cases. Model weights are promised "later," which will be the real test. Open weights invite independent benchmarking, red-teaming, and the kind of third-party validation that these self-reported numbers currently lack. Until then, the claims are plausible but unverified — impressive on paper, awaiting confirmation under real-world conditions.
