MiniMax M3.1 Flash: Cheaper AI Agents at 200 Tokens a Second

MiniMax has released M3.1 Flash, a faster and much cheaper update to the M3 model it launched in June, and it is aimed squarely at the economics of AI agents. On October 19, 2026 the company said agentic tasks cost 70% less to run than on its previous model, while online inference is nearly three times faster. It also solved a small community mystery: "Space Bunny," the anonymous model that processed more than 200 trillion tokens on OpenRouter, was an early build of M3.1 Flash.

• Price: $0.15 per million input tokens, $0.60 per million output tokens and $0.015 per million cached input tokens
• Speed: output above 200 tokens per second in everyday use, with median (P50) speeds above 170, according to MiniMax
• Availability: live now on the MiniMax API, the M Plan subscription and the MiniMax Code desktop app; weights promised "later" with no date
• Headline scores (MiniMax's own testing): 61.7 on OSWorld V2, 76.2 on BankerToolBench, 73.7% on DeepSWE v1.1 and 82% on CyberGym

Why Agent Costs Are the Real Bottleneck

Agents rarely make a single model call. A coding agent that reads a repository, writes a patch, runs tests and fixes what broke can easily make dozens of calls before it finishes one job. That multiplies every cent of per-token pricing, which is why "Flash" tier models have quietly become the workhorses of agent pipelines. MiniMax credits the new speed to "more aggressive sparsity" at inference time rather than a smaller model. SiliconANGLE, citing MiniMax's developer guide, reports the model still has about 428 billion parameters in total, accepts text, image and video input, and handles prompts up to 1 million tokens, with reasoning always on and five effort levels from Low to Max.

What Can MiniMax M3.1 Flash Actually Do?

MiniMax groups its claims into four areas, and all of the numbers below come from the company's own evaluations rather than independent testing:

• Complex office workflows: the company says M3.1 Flash ranks among the top performers on AutomationBench-AA, completing end-to-end tasks across SaaS tools in sales, finance, HR and customer service
• Coding: 73.7% on DeepSWE v1.1, a benchmark where the model must understand a codebase, build features, fix issues and verify its own work through tests
• Computer use: 61.7 on OSWorld V2 and 29.6 on Agents' Last Exam, which MiniMax says ranks first among Chinese models on both
• Cybersecurity: 82% on CyberGym, a Berkeley-built benchmark that checks whether an agent can reproduce real software vulnerabilities, plus internal expert reviews of code auditing and vulnerability patching

For context, SiliconANGLE notes that the price matches DeepSeek's off-peak rate for V4.1 Flash, which DeepSeek says scored 77% on the same DeepSWE test. Neither result has been independently verified yet, but the takeaway is clear: frontier-adjacent agent performance is getting dramatically cheaper. We covered that DeepSeek release in our look at DeepSeek V4.1 Flash 1017.

How Does the Price Compare to the Original M3?

MiniMax frames the headline saving at the task level: for agentic work, total usage cost lands 70% below the previous model. Some of that comes from the flat per-token price and some from finishing tasks faster with fewer wasted steps. The 1.5-cent cached-input rate matters too, because agents re-send the same system prompt, tool definitions and file context on almost every turn. If you have been prototyping on the original MiniMax M3 open-weight model, the upgrade path is mostly a model-name change in your API calls.

What Comes Next for MiniMax

Two things are still on the way. MiniMax says the Flash weights will be published later, which would put a fast, agent-tuned model within reach of self-hosters, and a more capable "Pro" model is also coming. Until the weights arrive, the practical move for builders is simple: rerun your own agent evaluations on M3.1 Flash and compare cost per completed task, not cost per token. That is the number that decides whether an agent is a demo or a product. For more model releases, browse our AI coverage.

Sources: MiniMax — MiniMax M3.1 Flash: Redefining Agent Performance and Cost — October 19, 2026; SiliconANGLE — MiniMax launches M3.1 Flash model — October 19, 2026.