MiniMax Releases M3.1 Flash, a Low Cost Agent Model With Open Weights Promised Later

MiniMax on October 19, 2026 released M3.1 Flash, a model for agent tasks, coding, computer use and cybersecurity, through its API, M Plan subscription and MiniMax Code app. The company says online inference is nearly three times faster and usage costs on agent tasks are 70% lower than with its predecessor.

MiniMax shipped M3 on June 1, 2026 and called it the first open weight model to combine frontier level coding, a 1 million token context window and native multimodal input. M3.1 Flash is a faster, cheaper follow up, and it lands two days after DeepSeek's V4.1 Flash 1017, making it the second low cost Chinese model release aimed at agent work in one week.

What did MiniMax release on October 19?

MiniMax released M3.1 Flash on October 19, 2026 as a "next-generation Flash model" and said in its announcement that model weights "will be released later." The Decoder reports that the open weight release is planned for October 23.

MiniMax says an early version of the model ran anonymously under the name Space Bunny and processed more than 200 trillion tokens before launch. The company credits the speed gain to what it calls "more aggressive sparsity" and claims output speeds above 200 tokens per second in everyday use, with a median of more than 170. MiniMax also says a Pro version will arrive later.

How does M3.1 Flash score against DeepSeek, Claude and GPT?

In MiniMax's own table of 15 benchmarks, published October 19, 2026, M3.1 Flash has the top score on 3 rows: Automation-Bench, LOCA-Bench and BankerToolBench, where it ties Claude Opus 5 at 76.2. It trails DeepSeek V4.1 Flash 1017 on 5 of the 6 rows where both have a score, leading only on Automation-Bench.

Counted over the table, M3.1 Flash beats GPT-5.6 Sol on 7 of 15 rows and Claude Opus 5 on 2 of the 14 rows where both have a score, with one tie. Like DeepSeek's table two days earlier, MiniMax compares against Claude Opus 5 and GPT-5.6 Sol, although its own cost charts include the newer Claude Opus 5.5, GPT-6 Astra and GPT-6.1 Sol.

The two labs' tables do not agree on shared rows. On Agents' Last Exam, MiniMax lists Claude Opus 5 at 32.2 and GPT-5.6 Sol at 30.6 and leaves DeepSeek's model blank, saying it uses only full set results from the official leaderboard. DeepSeek's October 17 table gave Opus 5 28.6, GPT-5.6 Sol 26.7 and its own model 35.4, a self reported score above M3.1 Flash's 29.6.

What does M3.1 Flash cost per task?

MiniMax prices M3.1 Flash at a flat $0.15 per million input tokens, $0.60 per million output tokens and $0.015 per million cached input tokens, as of October 19, 2026. Those per token rates are half of what OpenRouter lists for MiniMax M3 ($0.30 and $1.20) and half of DeepSeek's peak rates for V4.1 Flash 1017, which MiniMax's own chart footnote gives as $0.30 and $1.20.

MiniMax's cost charts put M3.1 Flash at 62.8% on Automation-Bench for under $0.05 per task, ahead of DeepSeek V4.1 Flash 1017 at 60.5% for roughly $0.09, according to the chart. On DeepSWE v1.1 the picture flips: M3.1 Flash reaches 73.7% at just under $0.30 per task, while DeepSeek's model scores 77.0% for about $0.20 in the same chart. On Agents' Last Exam, MiniMax estimates a total cost of $111 for its 29.6% run, against well over $1,000 for Claude Opus 5.5, which scores about 38% in the chart.

Independent numbers point the same way on cost. Artificial Analysis gives M3.1 Flash 42 points on its Intelligence Index, up from 29 for M3, and The Decoder reports a cost of about $0.22 per task on that test run against $0.26 for DeepSeek V4.1 Flash 1017, which scores 45. Artificial Analysis also measured 120.1 output tokens per second, below the median of more than 170 that MiniMax claims, and The Decoder notes a 93% hallucination rate on the AA-Omniscience test because the model rarely declines to answer.

Who can use M3.1 Flash, and where?

M3.1 Flash is available as of October 19, 2026 through the MiniMax API at platform.minimax.io, the M Plan subscription and the MiniMax Code desktop app, according to MiniMax. Third party platforms added it the same day.

Vercel added the model to its AI Gateway as minimax/minimax-m3.1-flash, free until October 22 when MiniMax itself serves the request, after which it bills at MiniMax's standard rate. OpenRouter lists a free variant served by MiniMax.

What we don't know yet

• Whether MiniMax publishes the weights on October 23, as The Decoder reports, and under which license.
• The model's parameter count and architecture details beyond the sparsity claim.
• How M3.1 Flash scores against Claude Opus 5.5, GPT-6 Astra and GPT-6.1 Sol on MiniMax's full benchmark set.
• When the Pro version that MiniMax mentions will launch.