MiniMax releases M3.1 Flash and claims 70% lower AI agent costs
MiniMax says its M3.1 Flash model is faster and cheaper than M3, but its agent-cost and benchmark claims have not been independently replicated.
By Colin Brandt · Enterprise Reporter
· 2 min read
MiniMax released M3.1 Flash on Oct. 19, pitching the update to its M3 model around speed and the MiniMax M3.1 Flash cost of running AI agents. The company says agentic workloads cost 70% less than on its previous model and that inference is nearly three times faster, claims that could matter for products that make repeated model and tool calls to complete a task.
Those are MiniMax’s figures, not independently validated savings. The supplied reporting does not describe a standardized workload, deployment setup or task-success methodology that would let customers apply the 70% figure to their own agents. A lower price per token also does not automatically mean a lower cost per completed task.
What does MiniMax M3.1 Flash cost for AI agents?
MiniMax lists API pricing of $0.15 per million input tokens, $0.60 per million output tokens and $0.015 per million cached input tokens, according to reporting from SiliconANGLE and Forkast News. SiliconANGLE reported that the input rate matches DeepSeek V4.1-Flash’s off-peak input pricing.
End-to-end agent bills depend on more than those rates: the input and output token mix, cache use, chosen reasoning level, number of model calls, tool execution and whether the model completes the task successfully all affect cost. MiniMax offers five reasoning-effort settings, from Low to Max, with reasoning enabled on each request. It reported output exceeding 200 tokens per second in everyday use, with median throughput above 170 tokens per second in other coverage.
The model has 428 billion total parameters, supports text, image and video inputs, and accepts prompts up to 1 million tokens, according to SiliconANGLE. The company attributes the speed increase to more aggressive sparsity during inference.
MiniMax’s reported benchmark results
MiniMax published results aimed at coding, computer use, finance-oriented tool work and cybersecurity. The reported scores are useful for comparing the company’s stated positioning, but they are not independent test results.
- OSWorld V2: 61.7
- BankerToolBench: 76.2
- DeepSWE v1.1: 73.7%
- CyberGym: 82%, with one report listing 82.2%
- Agents’ Last Exam: 29.6%
SiliconANGLE reported that MiniMax’s 73.7% DeepSWE score trailed DeepSeek’s reported 77% on the same software-engineering benchmark. The report said neither company result had been independently checked. Forkast News likewise reported that MiniMax had not published a technical paper or formal model card and that the vendor-provided performance measures had not been independently replicated.
For operators, M3.1 Flash is a pricing and throughput claim that needs workload-specific testing. MiniMax has supplied token rates and benchmark figures; the supplied reports do not establish universal 70% savings or validate the reported task performance outside the company’s evaluations.
This story draws on original reporting from SiliconANGLE.
