MiniMax M3.1 Flash focuses on speed, code, and agents
MiniMax launched M3.1 Flash on October 19, 2026, presenting this model as a generational leap for agent tasks, coding, and cybersecurity. The version had been quietly tested for several days under the name Space Bunny.
MiniMax officially launched M3.1 Flash on October 19, 2026, a model presented as a generational leap for agentic tasks, coding, computer use, and cybersecurity. Behind this launch, the company also confirms that Space Bunny, an anonymous model having processed over 200 trillion tokens in stealth mode, corresponded to an early version of this iteration. The message is clear: the battle is now as much about speed as cost, with very concrete uses for businesses. And, by the way, benchmarks were not invited just to decorate the press release.
Un lancement nourri par des tests à l’ombre
Space Bunny was not just a codename. According to OpenRouter, which hosted the trial, the model was number one for 10 consecutive days and processed 56.7 trillion tokens, a record for a stealth model on its platform. OpenRouter also states that the model left its service on October 5, 2026, before returning on October 19 under the name M3.1 Flash, with a free 72-hour offer. MiniMax says it used this feedback to fix bugs and improve the product before its public release.
The announced gains are measurable. OpenRouter reports a time to first token reduced from 1.95 seconds to 1.22 seconds, a 38% improvement, throughput increased from 70 to 100 tokens per second, and an error rate reduced from 1.4% to 0.7%. MiniMax, for its part, claims speeds exceeding 200 tokens per second in regular use, with a median above 170 tokens per second. In practice, this changes the user experience as much as integration costs.
Ce que MiniMax met derrière “agentique”
MiniMax describes M3.1 Flash as a model designed for complete workflows, not just to answer isolated questions. The company cites sales, marketing, operations, customer service, finance, and human resources tasks, executed end-to-end in SaaS environments. In other words, the model is designed to chain actions, keep track, and handle multiple tools without losing direction — what product teams like to call “agentique,” and what users mainly want to see work smoothly.
On AutomationBench-AA, MiniMax announces a score of 72.5% according to OpenRouter, and 76.2% according to its own release. On BankerToolBench, which measures multi-step planning, document consistency, financial compliance, and chaining of tool calls, the company claims 76.2%. For software development, it cites 73.67% on DeepSWE v1.1 in OpenRouter’s report, and 73.7% in its announcement. These presentation differences do not change the overall trend but justify a [to be verified] if comparing tables line by line.
Codage, ordinateur et cybersécurité : le trio que tout le monde vise
The model’s positioning goes beyond a simple chat assistant. MiniMax says M3.1 Flash ranks among the best on complex agent tasks, computer use, software development, and cybersecurity. On OSWorld V2, a benchmark evaluating the ability to operate a computer and chain tasks between applications, the company announces 61.7 points. On Agents’ Last Exam, it cites 29.6 points and claims to lead among Chinese models on these two measures.
Cybersecurity holds a central place in this launch. MiniMax reports 82% on CyberGym, which it presents as a vulnerability discovery and validation test. The company adds that its internal experts observed good performance in code auditing, vulnerability fixing, penetration testing, malware analysis, and reverse engineering. This is not just a matter of technical prestige: a model capable of following complex action chains interests both security teams and agent integrators.
Prix serré, accès large, ambition nette
The announced pricing is aggressive. MiniMax lists $0.15 per million input tokens, $0.60 per million output tokens, and $0.015 per million cached input tokens. The company also says the overall usage cost for agentic tasks is 70% lower than the previous model, with a promise of “frontier” level at one-third the cost of M3 and three times the speed. These figures, if maintained in real use, provide a serious argument against well-established competitors.
M3.1 Flash is available via the MiniMax API, the M Plan subscription, and MiniMax Code for desktop. OpenRouter announces the same pricing grid after its free 72-hour period. MiniMax finally specifies that the model weights should arrive on October 23, 2026, while a Pro model, presented as even more ambitious, will come later. In this context, Flash looks less like a lightweight version and more like a working base to test how far the balance between efficiency and capabilities can be pushed. The market, meanwhile, is mainly watching to see if the promise holds beyond the press release.