MiniMax M3.1 Flash Took On Open Models In Our Newsroom Test

🔍 Read the full analysis: MiniMax M3.1 Flash Took On Open Models In Our Newsroom Test on ThorstenMeyerAI.com

Before you orderOffer from Amazon

Get the little things that make your day delivered free with Prime

• Fast, free delivery on millions of items
• Prime Video, Amazon Music and more included
• Member-only deals all year

Start your free Prime trial →Free trial for eligible customers · Cancel anytime

As an affiliate, we earn on qualifying purchases.Hide

MiniMax released M3.1 Flash on October 19, pricing the model at $0.15 per million input tokens and $0.60 per million output tokens, and promoting it for agent work. A report by ThorstenMeyerAI.com compared MiniMax’s claims with Artificial Analysis measurements and a newsroom test on 24 article assignments; the supplied account describes the test design but does not provide its scores, leaving that part of the comparison unresolved.

MiniMax says M3.1 Flash is about half the price per token of M3 and costs about 70% less per agent task. Those savings are company claims, not results established by the newsroom test details provided. MiniMax also says the model can exceed 200 tokens per second in everyday use, with a median above 170, and is nearly three times faster than M3. It attributes the change to “more aggressive sparsity” but has not published further design details in the source account.

The model has a one-million-token context window, adjustable reasoning effort and access through the MiniMax API, M Plan subscription, MiniMax Code and OpenRouter. OpenRouter listed it free for its first 72 hours. The stated paid rates are $0.15 per million input tokens, $0.60 per million output tokens and $0.015 per million cached input tokens. MiniMax says the weights will arrive October 23, but the release had not happened when the report was published.

Artificial Analysis gave M3.1 Flash an Intelligence Index score of 43. In the comparison reproduced by the report, DeepSeek V4.1 Flash 1017 scored 45, as did GLM-5.3; several other open-weight models scored 44 or 46. Artificial Analysis measured M3.1 Flash at 120 output tokens per second and 1.08 seconds to first token, below MiniMax’s speed claim. Its index run used 180 million tokens, compared with a reported median of 99 million. A larger token count can affect real costs even when the per-token price is low.

Price Gains Meet Usage Costs

M3.1 Flash’s low listed rates could appeal to developers running high-volume agent workloads, where token costs accumulate quickly. But price per token does not alone establish the cost of completing a task. Artificial Analysis found the model used more tokens than the index’s median, while the company’s claim of roughly 70% lower cost per agent task was not independently verified in the supplied report.

The mixed measurements also matter to teams choosing models for coding or automation. MiniMax’s published table gives M3.1 Flash a lead on Automation-Bench, but DeepSeek V4.1 Flash 1017 scores higher on DeepSWE and Terminal-Bench. The comparisons are limited: MiniMax selected the competitors, some newer rivals were not included, and the supplied newsroom account does not state its model scores. Readers therefore have evidence about price and selected benchmark results, not a complete picture of performance in production.

For newsrooms, the practical question is whether a model can produce accurate copy from source material without adding unsupported details. The test was designed around that concern, with two judges scoring fidelity and invented facts alongside clarity, usefulness, structure and polish. Without the scores, no conclusion about M3.1 Flash’s editorial reliability can be drawn from the material provided.

From Space Bunny to M3.1

Before its public identification, the model ran anonymously on OpenRouter as Space Bunny Alpha. According to the report, it appeared on September 23, led OpenRouter’s usage ranking for 10 days and processed 56.7 trillion tokens on the platform before being removed on October 5. MiniMax says its overall usage exceeded 200 trillion tokens. These figures describe different scopes and are attributed to the report and company, respectively.

The report says that during the anonymous period the provider could retain prompts but was not permitted to train on them. MiniMax has not said what happened to that data. M3.1 Flash’s weights were also still pending, so Artificial Analysis listed it as proprietary at the time. The report says MiniMax had not published a license or official parameter count; an outside estimate cited elsewhere was not confirmed by the company and was excluded from the analysis.

MiniMax’s own benchmark chart reports a 62.2 score for M3.1 Flash on Automation-Bench, compared with 60.5 for DeepSeek V4.1 Flash 1017. DeepSeek scores higher on DeepSWE, 77.0 to 73.7, and Terminal-Bench, 42.3 to 38.1. The report cautions that the chart is company-selected and self-reported. It also says MiniMax’s claim that the model ranks first among Chinese models on Agents’ Last Exam omits DeepSeek from that row, while DeepSeek’s model card reports a different score that may use another setup.

““more aggressive sparsity””

— MiniMax

Scores and Data Questions Remain

The supplied source material describes the newsroom test but ends before its results table. It says 12 source items were each turned into English and German articles, for 24 outputs, and that GPT-6.1 Sol and Claude Opus 5.5 judged them on a five-point scale. It does not provide M3.1 Flash’s scores or the comparison models’ results. The model’s newsroom performance therefore cannot be reported from the information available here.

Other open questions include whether the October 23 weight release occurred as scheduled, what license will govern the weights, and what the model’s parameter count is. MiniMax’s speed and agent-cost figures remain company claims alongside the lower speed measured by Artificial Analysis. The report also does not resolve how prompts retained during the Space Bunny period were handled.

Benchmark comparisons require caution. MiniMax chose the models and tests in its chart, some rival versions were newer than the listed comparisons, and the company’s figures had not been independently reproduced according to the report. Artificial Analysis provides a separate standardized view, but its index is not a substitute for the missing newsroom-test results or for testing a model on an individual organization’s workload.

Watch for Weights and Test Results

The next stated milestone is MiniMax’s promised October 23 weight release. Once the weights and license are available, developers can assess how they may be deployed and whether independent evaluations can reproduce the company’s benchmark claims. The source does not confirm that the release has occurred.

The newsroom comparison also needs its missing results table before readers can judge how M3.1 Flash performed against other models on the 24 article assignments. Until those scores are published, the evidence supports a narrower conclusion: the model launched at low token prices, its independent index results trail several listed open-weight competitors, and its editorial accuracy in this test remains unreported in the supplied material.

Key Questions

When did MiniMax release M3.1 Flash?

MiniMax announced and released M3.1 Flash on October 19, 2026, through its API and other listed services. The report says OpenRouter made it free for the first 72 hours.

How much does M3.1 Flash cost?

The listed rates are $0.15 per million input tokens, $0.60 per million output tokens and $0.015 per million cached input tokens. MiniMax also claims about 70% lower cost per agent task than M3, but the report does not establish that claim independently.

How did M3.1 Flash perform in the newsroom test?

The source describes a test of 24 articles in English and German, judged by GPT-6.1 Sol and Claude Opus 5.5. The supplied material does not include the results, so it does not establish how the model scored or whether it outperformed other models.

Are M3.1 Flash’s weights available?

At the time covered by the report, the weights had not been released. MiniMax said they would arrive on October 23, 2026. The report also says the company had not disclosed a license or official parameter count.

Is M3.1 Flash faster than competing models?

MiniMax claims everyday speeds above 200 tokens per second and a median above 170. Artificial Analysis measured 120 tokens per second and 1.08 seconds to first token, below the company’s stated speed. These measurements use different sources and should not be treated as equivalent without further test details.

Source: ThorstenMeyerAI.com