MiniMax M3.1 Flash halves token prices, 70 percent agent saving disputed

A per-million-token price tells you one piece of what an agent task costs; the other piece is how many tokens the agent spends to finish the job. MiniMax says its new M3.1 Flash is 70 percent cheaper per agent task than the model it replaces, which is a bigger saving than halving the per-token price alone would produce.

MiniMax released M3.1 Flash on Monday. Its own announcement puts the per-million-token prices at $0.15 for input, $0.60 for output and $0.015 for cached input, and claims inference is nearly three times faster than its predecessor. The company lists Space Bunny, which processed more than 200 trillion tokens anonymously before the launch, as an early version of this model.

What two independent numbers say

The New Stack points out that MiniMax has not said which workloads produced the 70 percent agent-cost figure. Artificial Analysis ran its own comparison, and in its announcement puts the average cost per Intelligence Index task at $0.22 for M3.1 Flash, down from $0.51 for M3. That is a 57 percent reduction, below MiniMax's own 70 percent claim, and the two figures come from different workloads.

The gap matters because a cheaper rate card only helps if the agent also finishes in fewer loops. The New Stack cites a study of token use in agentic coding tasks finding that repeated runs of the same task can use up to 30 times as many tokens, with input tokens driving most of the usage. For a team in New York or San Francisco working in daytime hours, DeepSeek's V4.1 Flash matches MiniMax's $0.15 input and $0.60 output rates off-peak, with a cheaper $0.003 cached input against MiniMax's $0.015.

Coding, computer use and the hallucination number

On coding, M3.1 Flash scores 73.7 percent on DeepSWE v1.1, which MiniMax lists against the 74 percent it reports for Claude Opus 5. On Artificial Analysis's Intelligence Index it scores 42, tied with GLM-5.3 Flash, ahead of Gemini 3.8 Flash at 41, and behind DeepSeek V4.1 Flash 1017 at 45. On computer use, OSWorld V2 gives it 61.7, which The New Stack notes trails the strongest Anthropic and OpenAI models.

Artificial Analysis's AA-Omniscience benchmark is where the model looks weakest. It answered 56 percent of factual questions correctly, ahead of DeepSeek V4.1 Flash at 43 percent, but attempted nearly every question, and about 93 percent of its non-correct responses were wrong answers rather than abstentions. The summary from Artificial Analysis is high accuracy offset by a high hallucination rate.

M3.1 Flash is available through MiniMax's API, M Plan and MiniMax Code. Weights are coming later, with no license yet published, so holding off on self-hosting plans is prudent until that document exists.

Source

MiniMax just halved its token prices. Your AI agent bill may not follow., by Amanda Caswell at The New Stack. Primary source: MiniMax M3.1 Flash: Redefining Agent Performance and Cost.