Inside Our Newsroom Test Of MiniMax M3.1 Flash

🔍 Read the full analysis: Inside Our Newsroom Test Of MiniMax M3.1 Flash on ThorstenMeyerAI.com

Buying for a business?Offer from Amazon

Get business pricing on office and shipping supplies

• Business-only prices and quantity discounts
• Tax-exempt purchasing
• Multiple users, one account, clear invoices

Create a free Business account →Free to register

As an affiliate, we earn on qualifying purchases.Hide

MiniMax released M3.1 Flash on October 19, pricing it at $0.15 per million input tokens and $0.60 per million output tokens and promoting it as a fast, lower-cost model for agent work. A review published by ThorstenMeyerAI.com compared MiniMax’s claims with an independent index and a 24-article newsroom test; the measurements present a more mixed picture, and the supplied source does not include the newsroom test’s scores.

The model is available through the MiniMax API, M Plan subscription, MiniMax Code and OpenRouter. OpenRouter offered it free for its first 72 hours, according to the review. Cached input is priced at $0.015 per million tokens. The model supports a one-million-token context window, adjustable reasoning effort and, on OpenRouter, an output limit of up to 524,288 tokens.

MiniMax says M3.1 Flash costs half as much per token as its predecessor, M3, and about 70% less per agent task. The company also claims everyday output speed above 200 tokens per second, with a median above 170, and says it is nearly three times faster than M3. Those are company claims: the design description provided is limited to “more aggressive sparsity.”

Artificial Analysis measured the model at 43 on its Intelligence Index, 120 output tokens per second and 1.08 seconds to the first token. The index’s listed open-weight leaders scored 44 to 46. The review also reports that M3.1 Flash used 180 million tokens to complete the index, compared with a 99-million-token median. That is a token-use comparison for the evaluation, not a direct measure of the cost of every user task.

Lower Prices, Mixed Performance

The release puts a specific price and performance proposition in front of developers: lower per-token costs paired with agent-oriented features and a large context window. For teams running repeated model calls, the bill depends not only on the listed token price but also on how many tokens a model uses and how well it completes the task. Artificial Analysis’s higher-than-median token use is relevant to that calculation, though it does not by itself establish real-world costs across different applications.

The independent index also complicates a simple speed-and-price pitch. Its measured 120 tokens per second is below MiniMax’s claim of more than 200 and below the 211 tokens per second listed for DeepSeek V4.1 Flash 1017. Meanwhile, MiniMax’s own benchmark chart shows M3.1 Flash leading on Automation-Bench but trailing DeepSeek on DeepSWE and Terminal-Bench. The differing results make task-specific testing more useful than relying on a single overall ranking.

For newsroom and publishing work, the review’s test design addresses a practical concern: whether a model can produce a usable article from source material without adding unsupported details. It used real news sources and two language versions, but the provided report excerpt gives no scores or model ranking. Its findings cannot be inferred from the setup alone.

From Space Bunny to Release

The model’s public release followed an anonymous period on OpenRouter. According to ThorstenMeyerAI.com, a model identified as Space Bunny Alpha appeared there on September 23, led the platform’s usage ranking for 10 days and processed 56.7 trillion tokens before being removed on October 5. MiniMax later said it had processed more than 200 trillion tokens overall; the source does not specify the period or provide a breakdown for that larger figure.

MiniMax identified Space Bunny Alpha as an early version of M3.1 Flash when it announced the model on October 19. The weights had not yet been released at the time covered by the review. MiniMax said they would arrive on October 23, meaning Artificial Analysis listed the model as proprietary for its index. There was no official parameter count or license in the supplied material.

The review also notes that MiniMax selected the competitors in its published benchmark comparison, and that several named rivals were older versions. Its benchmark figures were transcribed from company chart images. The comparison is therefore informative as a record of what MiniMax reported, but it is not an independent, like-for-like evaluation of every current competing model.

““more aggressive sparsity””

— MiniMax

Scores and Data Still Missing

The supplied source ends before presenting the newsroom benchmark results. It describes 12 news sources, each written in English and German, scored by GPT-6.1 Sol and Claude Opus 5.5 for source fidelity, invented facts, clarity, usefulness, structure and polish. But it gives no model scores, comparison table, or conclusion. The test’s outcome and the model’s relative performance on newsroom tasks therefore remain unknown from the material provided.

Other points are unresolved. MiniMax has not confirmed an official parameter count or announced a license in the source. A figure of 428 billion total and 23 billion active parameters was attributed to another outlet, but the review says it is unconfirmed. MiniMax’s reported benchmark and security results, including an 82% CyberGym score, have not been independently reproduced in the account.

The review also flags a data-handling question from the anonymous OpenRouter period: the provider could retain prompts, though not use them for training. MiniMax had not said what happened to that data. The report does not establish whether any particular user’s prompts were retained or how long they may have been kept.

Weights and Further Testing

The next stated milestone is the promised October 23 release of the weights. If they are released as MiniMax says, the license and model files will help clarify whether and how developers can run or adapt the model outside the hosted services. The source provides no confirmation that the release has happened.

Further scrutiny will depend on independent, reproducible evaluations, clearer disclosures about the model’s size and licensing, and clarification of prompt handling during the Space Bunny Alpha period. For readers considering the model, the review supports comparing performance and token use on the tasks they actually run rather than treating the company’s price, speed or benchmark claims as settled results.

Key Questions

When did MiniMax release M3.1 Flash?

MiniMax released M3.1 Flash on October 19, 2026, according to the source report.

How much does M3.1 Flash cost?

The listed price is $0.15 per million input tokens and $0.60 per million output tokens. Cached input is listed at $0.015 per million tokens.

Did the newsroom test find that M3.1 Flash writes accurate articles?

The supplied report describes a test of 24 articles, scored by two judges, but does not include the scores or findings. No conclusion about its newsroom accuracy can be drawn from the provided material.

Are M3.1 Flash’s weights available?

They had not been released at the time covered by the review. MiniMax said the weights would arrive on October 23; the supplied source does not confirm whether they were subsequently published.

Source: ThorstenMeyerAI.com