MiniMax M3.1 Flash: faster AI agents at 70% lower cost?

MiniMax M3.1 Flash brings a 1M-token context and OpenAI- and Anthropic-compatible APIs. We assess its speed, pricing, benchmarks and caveats.

Main report

MiniMax officially released M3.1 Flash on October 19, 2026 through its API, M Plan and MiniMax Code. The company positions it as a multimodal model for AI agents, coding, tool use and long-context work. Reuters independently confirmed the launch and MiniMax's two headline claims: online inference is nearly three times faster and agent-task costs are 70% lower than its predecessor.

What has actually been published

MiniMax's API documentation lists a maximum context window of 1,000,000 tokens, text, image and video input, and five reasoning-effort levels from 'low' to 'max'. Posted prices are $0.15 per million input tokens, $0.60 per million output tokens and $0.015 per million cached-input tokens. OpenAI- and Anthropic-compatible endpoints can reduce integration work.

MiniMax says output can exceed 200 tokens per second in everyday use, with median P50 speed above 170 tokens per second. Those measurements come from the vendor's infrastructure; real latency will vary with server location, load, context length and effort level.

Analysis

What the benchmarks do—and do not—show

Vendor-reported results include 73.7% on DeepSWE v1.1, 61.7 on OSWorld V2, 76.2 on BankerToolBench and 82% on CyberGym. The pattern suggests optimization for multi-step workflows, tool calls and code rather than chat alone.

Provider benchmarks are not a substitute for independent evaluation. Teams should test completion rate, human correction, cost per successful job, P95 latency and recovery from tool failures on their own workloads. The 70% cost reduction is MiniMax's comparison with its predecessor on agentic tasks; it does not guarantee that every application will save exactly 70%.

M2.5 and M2.7 stopped accepting new users on October 19 and are scheduled for retirement on November 18, 2026. M Plan and Token Plan requests may be routed automatically to M3.1 Flash, while pay-as-you-go applications should compare outputs before completing a migration.

What to watch

A practical way to evaluate M3.1 Flash

Start with one narrow, measurable workflow such as support-ticket classification, code review backed by tests or internal-document synthesis. Run it beside the current model on the same task set and record quality, elapsed time, token use, tool-call failures and human intervention. Before sending sensitive data, review retention, processing-location and access-control terms.

A one-million-token window can help with large repositories or document sets, but filling it is not automatically efficient. Targeted retrieval, caching and an appropriate reasoning level may influence real cost more than the list price. MiniMax says model weights will be released later, without a specific date, so self-hosting should not yet be treated as available.

Read Mistral Large 4 enters API preview for another model-release comparison.

Verified sources: MiniMax's October 19 announcement, MiniMax API documentation, the model migration notice and Reuters, October 19, 2026.