MiniMax M3.1 Flash Cuts Agentic Costs

The anonymous Space Bunny model that chewed through more than 200 trillion tokens, MiniMax says, was an early version of M3.1 Flash. That reveal matters less as trivia than as a clue that the company has been training and serving this stack in the wild before branding it.

In brief

MiniMax (19 October 2026) has rolled out M3.1 Flash, saying it delivers a “generational leap” in agentic tasks, coding, computer use and cybersecurity while making online inference nearly three times faster and cutting usage costs by 70%.
For teams running long tool-using jobs, the pitch is cheaper tokens and higher throughput: $0.15 / $0.60 per million input/output tokens, with claimed output speeds of 200+ TPS.
What’s still open: model weights come later — this is an API launch for now — and every benchmark score is MiniMax’s own.

A Flash model aimed at agents, not just chat

MiniMax frames M3.1 Flash as a next-generation Flash model built for multi-step work: enterprise SaaS workflows, software engineering loops, desktop computer use and security tasks. On its own numbers it scores 73.7% on DeepSWE v1.1, 61.7 on OSWorld V2, 29.6 on Agents’ Last Exam, 76.2 on BankerToolBench and 82% on CyberGym, and says it ranks first among Chinese models on OSWorld V2 and ALE.

Metric (MiniMax)	Figure
Input / output / cached	$0.15 / $0.60 / $0.015 per 1M tokens
Cost vs predecessor	70% lower for agentic tasks
Output speed	200+ TPS; P50 170+ TPS
Weights	Later (API, M Plan, MiniMax Code now)

Source: MiniMax blog, 19 October 2026. Scores and speed claims are the company’s.

Those figures will face the usual independent-eval test. Until then they sit as vendor claims, useful for pricing conversations but not as a league table. The company also says everyday output can hit 200+ tokens per second, with median speeds above 170 TPS, after more aggressive sparsity work.

Available now, open weights later

M3.1 Flash is live through the MiniMax API, M Plan and MiniMax Code. A Pro variant “will also arrive later,” the company says, for teams that want more raw intelligence once the cost-focused Flash line is shipping. That sequencing keeps the weight-release promise separate from today’s API launch — and means nobody should treat M3.1 Flash as an open-weight drop yet.

That pricing table is the commercial hook. At $0.15 input and $0.60 output per million tokens, with cached input at $0.015, MiniMax is pitching M3.1 Flash as a volume workhorse for agent loops that burn tokens on tool calls and retries. The company says overall agentic usage cost is 70% lower than its predecessor — again a MiniMax comparison, not a third-party audit.

For developers, availability through MiniMax Code and M Plan sits alongside the raw API, so the same model can show up in a desktop agent product and in custom backends. The delayed weight release still defines the story’s unfinished business.

Our read

Cheap, fast agentic models are becoming the default battleground after every “Flash” release from the big labs. MiniMax’s move is to own the cost-per-successful-workflow story and to admit Space Bunny was a soft launch. The kicker is the weights: without them this stays a closed API bet, and the next story is whether the delayed open release lands before rivals reset the price floor again.