MiniMax M3.1 Flash Launches with 70% Lower Agentic AI Costs

MiniMax launched the M3.1 Flash model on October 19, 2026, positioning it as a direct successor to the June 2026 M3 release. The new model targets agentic AI, coding, computer use, and cybersecurity workloads with vendor-reported improvements in speed and cost efficiency. Early versions operated under the codename Space Bunny on platforms such as OpenRouter, processing more than 200 trillion tokens before the public announcement.

MiniMax M3.1 Flash Pricing and Speed Claims

The API uses flat pricing of $0.15 per million input tokens, $0.60 per million output tokens, and $0.015 per million cached input tokens. MiniMax states that overall usage costs for agentic tasks drop by 70 percent compared with the prior M3 model. Online inference runs nearly three times faster, with median output speeds above 170 tokens per second and peaks exceeding 200 tokens per second through increased sparsity.

Technical specifications include a 1-million-token context window, support for text, image, and video inputs with text output, and always-on reasoning across five effort levels from Low to Max. The architecture uses a mixture-of-experts design with approximately 428 billion total parameters and roughly 23 billion active parameters per token. Independent evaluations placed the model at 42 on the Artificial Analysis Intelligence Index, an increase from 29 on the earlier M3 version.

Company benchmarks report 73.7 percent on DeepSWE v1.1, 61.7 percent on OSWorld V2, 29.6 percent on Agents’ Last Exam, 76.2 percent on BankerToolBench, and 82 percent on CyberGym. These scores reflect performance on enterprise automation and security tasks, though one independent test noted a high hallucination rate when the model attempts most questions.

Availability, Background, and Next Steps

Access is available immediately through the MiniMax API at platform.minimax.io, the M Plan subscription tier, and the MiniMax Code desktop application. Limited free tiers appear on certain third-party platforms. Model weights remain closed at launch but are scheduled for release on October 23, 2026, according to multiple reports.

The model builds on prior stealth testing that began around September 23, 2026. MiniMax has indicated that a higher-capability Pro variant will follow later. Users can begin testing cost and latency claims through the current API endpoints while waiting for open weights and additional third-party agent benchmarks.

Enterprise teams evaluating LLM pricing for agentic workflows can compare these rates directly against existing providers by running identical task suites. Early adopters should monitor real-world token consumption and hallucination rates on their specific automation and cybersecurity use cases before committing to larger deployments.