MiniMax cuts M3.1 Flash token prices in half, but agent bills may not follow

MiniMax released M3.1 Flash on Monday, saying the model cuts agent costs by 70 percent. Per-token prices fell by half, which explains part of the claim. The rest depends on how many tokens an agent burns to finish a job, and MiniMax has not said how it measured that.

The company lists $0.15 per million input tokens, $0.60 per million output tokens and $0.015 per million cached input tokens, and claims inference nearly three times faster than its predecessor, M3. It also supports a one-million-token context window, enough to fit a large codebase into a single prompt. Before launch, the model ran anonymously as “Space Bunny” and processed more than 200 trillion tokens.

Independent evaluation offers a narrower number. Artificial Analysis puts the average cost per Intelligence Index task at $0.22 for M3.1 Flash, down from $0.51 for M3, a 57 percent reduction. The two figures come from different workloads, so neither settles the argument.

Coding scores barely separate. M3.1 Flash lands at 73.7 percent on DeepSWE v1.1, a hair behind the 74 percent MiniMax lists for Claude Opus 5 and a point ahead of GPT-5.6 Sol. On Artificial Analysis’s Intelligence Index it scores 42, tied with GLM-5.3 Flash. Computer use is the weak spot: 61.7 on OSWorld V2, which MiniMax says leads Chinese models but trails the strongest Anthropic and OpenAI systems.

The more consequential finding is about reliability. On Artificial Analysis’s AA-Omniscience benchmark, M3.1 Flash answered 56 percent of factual questions correctly, ahead of DeepSeek V4.1 Flash at 43 percent. But it attempted nearly every question, and about 93 percent of its non-correct responses were wrong answers rather than abstentions.

The price comparison is not one-sided either. DeepSeek V4.1 Flash matches the $0.15 input and $0.60 output rates off-peak, doubles them during weekday peak hours, and charges $0.003 per million cached tokens against MiniMax’s $0.015. Since input tokens drive most agent costs, the cache rate may matter more than the headline price.

MiniMax also reports 82 percent on CyberGym and says internal security experts rated the model strong at code auditing, patching, penetration testing and reverse engineering. The launch post does not explain how those evaluations were run, so the security claims remain the company’s own until someone independent tests them.

The weights are still missing. MiniMax says they are coming “later” without publishing a license, and VentureBeat reports they may arrive later this week. A stronger Pro model is also in the works with no date. For teams that want to run the model on their own hardware, the price cut is only the first question.