On October 19, 2026, MiniMax officially released the M3.1 Flash model, a new iteration of the Flash line optimized for agentic tasks, coding, and Computer Use. According to the company, online inference became nearly three times faster, and agentic scenarios became approximately three times cheaper compared to M3. The model is already publicly available via API, subscription, and desktop application, while open weights are reportedly set to appear on October 23.

image
image
image

What happened

The company confirmed that the anonymous model "Space Bunny," which topped global rankings, was an early version of M3.1 Flash: before its name was revealed, it had already processed over 200 trillion tokens in real-world scenarios. The API pricing is as follows: $0.15 per 1 million input tokens, $0.60 per 1 million output tokens, and $0.015 for cache reads; according to vendor data, online inference speed exceeds 200 tokens per second. According to a MiniMax report, the model scores 62% on Automation-Bench, 73.7% on DeepSWE v1.1, 82% on CyberGym, 91% on BabyVision, and 76.2 on BankerToolBench. A weak result was also noted: on Terminal Bench 4.0, the model scored 38.1 compared to 51.8 for Claude Opus 5.

Context

The focus of this iteration is not universal capability growth, but inference efficiency and price: the Flash line has historically been responsible for fast and cheap scenarios. The specialization is also indicated by the single disclosed pair with the flagship—strong agentic results against one notable failure suggest profile-specific optimization rather than a general quality leap. A second important layer is the launch method: before the announcement, the model operated anonymously on real production traffic, which reduces the likelihood of paper-drawn metrics, although it does not replace independent quality assessments. The release falls into the segment of mass automation and coding agents, where model choice is determined by cost per task rather than peak accuracy, and where the upper price limit is held by closed flagships.

Why this matters for the industry

For the industry, this is primarily a shift in unit economics: with the new pricing, the cost of a typical agentic task is recalculated downward, the margins of existing services grow without code changes, and directions that previously did not add up financially become economically viable—mass document processing, multi-step automation, and large-scale code review. If independent runs confirm at least some of the claimed metrics, a routing template of "cheap model for mass tasks plus escalation to a flagship for the terminal-repository tail" will become established in product architectures. In the segment of mass automation and coding agents, this increases pressure on closed flagships like Claude Opus 5 and GPT-5.6: the premium for peak accuracy is no longer automatically paid when a comparable level in agentic workloads is available significantly cheaper.

Why this matters for users

The model can be tried immediately: via API on platform.minimax.io, in the M Plan subscription, and in the MiniMax Code desktop application, while on OpenRouter it is free for the first 72 hours from October 19. A practical first step is to run your own agentic pipelines on real tasks, calculate the actual cost per task at the new pricing, and compare quality with your current model on your own test set. According to independent reviews, the model's strength is office automation, while for complex terminal and repository tasks, it is advisable to keep a fallback to a flagship model. After the publication of open weights, the model can be deployed independently.

What is currently unknown / limitations

For now, this is a vendor-side picture: benchmark figures were reported by MiniMax itself without methodology, and there are no independent runs at the time of publication. The statement about a 70% reduction in agentic task costs is an aggregated estimate, dependent on the workload composition: the share of cache reads, the volume of output tokens, and the number of steps; the hard fact is only the price per token. The conditions for measuring speed and price (batching, caching, hardware) have not been disclosed. Open weights have not yet been published, and metrics can be directly verified only after October 23, while the comparison with flagships currently relies on a single pair of results.

Sources

Author

Look at AI, editorial team