MiniMax M3.1 Flash first look: careful with product data, unreliable on long jobs

MiniMax released M3.1 Flash on 19 October and said it matches frontier models on agent, coding and visual tasks at a third of the cost of MiniMax M3. It also said the anonymous "Space Bunny" model that has been busy on OpenRouter was an earlier version of it.

We ran six of our benchmarks on it the same evening. It writes some of the most reliable product listings of any cheap model, and it adds far fewer made-up claims than DeepSeek's new Flash. But on long jobs it often thinks until it runs out of room and returns nothing.

MiniMax M3.1 Flash at a glance

• Use it for product data and short copy at volume, where it was 9th of 24 on our listing test and first among models at its price.
• Don't use it for long analysis or report writing without a fallback: half our survey tasks and a quarter of our slide decks came back empty.
• Price: $0.15/$0.60 per million input/output tokens, flat. Weights promised for 23 October.
• Speed: Artificial Analysis measures 175 tokens a second on MiniMax's API. We could only reach it through a free, rate-limited route, so our own timings don't count.

What MiniMax says

MiniMax's announcement claims 62% on Automation-Bench, 30% on Agents' Last Exam, 74% on DeepSWE and 91% on BabyVision, "at 1/3 the cost of M3 and 3x the speed". The flat price is $0.15 per million input tokens and $0.60 per million output, with cached input at $0.015, and the weights are due on 23 October. Those are the developer's claims.

MiniMax M3.1 Flash benchmarks, measured by Spring Prompt

The model page has every result as it lands. How to read this table: each row is one benchmark, with the model's score, its rank among the models we have tested on it, and the best model in each group.

Scroll sideways for all 5 columns →

Ranked on the release we publish with this post. A group only counts models that benchmark has measured. GPT-6 Luna and Claude Haiku 5.5 ($0.20 blended) fall just below the price group; see the alternatives. BulletBench is not in the table: MiniMax lost its first four games on time at every clock and setting, on a rate-limited route.

Product data: careful, and cheap

CatalogBench gives a model 56 invented products, with source data and images, and asks for a listing that could go live with no human fix, three times over. With a factual brief, MiniMax passed 63.1% of listings and was right in all three repeats for 50.0% of products, 9th of 24 and just behind GPT-6 Luna (55.4%). It added about 22 unsupported claims per run, against 95 for DeepSeek-V4.1-Flash-1017 and 30 for the September V4.1 Flash.

The sales brief is where it stands out. With a brief to sell, most models start inventing benefits; MiniMax added 34 unsupported claims per run and was reliably right on 37.5% of products, 6th of 24. Only the GPT-6 models and Muse Wonder did better. One product per run came back empty.

Long jobs: it runs out of room

On SurveyBench, which asks a model to plan a survey and then analyse supplied responses, MiniMax used its whole 32,000-token output budget thinking on 18 of 24 steps and returned nothing. Every model gets the same budget, and DeepSeek-V4.1-Flash-1017 finished every step within it. Half the tasks got no answer at all, so its rating is last of 12. On DeckBench two of its eight decks never arrived and none of the other six was ready to present; head to head it still rated 1,094, 14th of 23, close behind DeepSeek's new Flash.

Its first ROASBench run stopped in month 3 for the same reason; we set it aside and the one rerun scored 40.3, 12th of 23, level with DeepSeek-V4.1-Flash-1017. ROASBench is one simulated year per model, not evidence of real advertising returns.

In the playground

All 28 of our playground prompts ran on the free route too. These are one-off tests, not a benchmark. Its spinning hexagon kept the ball inside even at the top spin setting, which DeepSeek's new Flash failed, and its chess game gets every rule right down to stalemate. Three answers never drew anything: its cassette recorder sits on "Warming up the transport" because its code uses a material it never defines, and two of the craft scenes crash on load.

Third-party results

• Artificial Analysis (our fetch, 19 October): 42.4 on the Intelligence Index, against 45.1 for DeepSeek-V4.1-Flash-1017 and 43.4 for Claude Haiku 5.5, at 175 tokens a second on MiniMax's API. Its model page puts the cost of running the index at $0.22 a task, against DeepSeek's $0.26, and calls the model very verbose. Data sourced from Artificial Analysis.
• Arena (reported 19 October): 12th among open models on Code Arena's WebDev board, 32nd overall with 1,588.

Pending: Arena's text and factuality ratings, tau2-bench and Vectara's hallucination leaderboard. Until Arena rates it, its overall score on our leaderboard is provisional.

What people are saying

A selection, with how we read each one:

MiniMax M3.1 Flash vs DeepSeek-V4.1-Flash-1017, GPT-6 Luna and Claude Haiku 5.5

These are the cheap models people will weigh it against, including DeepSeek's new Flash from two days earlier. Prices are blended per million tokens (three input to one output) at list price.

Scroll sideways for all 7 columns →

On product data MiniMax is far more careful than DeepSeek's new Flash at half its price, and well ahead of Claude Haiku 5.5. GPT-6 Luna is still ahead of it on every row here, costs a little less, and finished every task, so it remains the cheap default for business writing. Haiku 5.5 is the safer cheap choice for long analysis.

Is MiniMax M3.1 Flash good? Who should use it

• Product data and short copy at volume, if you want a cheap model that mostly sticks to what you give it. Check the gaps: about one listing in 56 came back empty per run.
• Not for long analysis, reports or decks without a retry or fallback model. It often thinks until it runs out of room.
• Not for anything that must answer in a second or two, at least on the free route.
• Wait for the paid route and the weights before switching production work; we will rerun our tests on the paid route.

These benchmarks aren't your data, so test it on your own prompts first. Its 28 playground answers sit beside other models' answers to the same prompts.

What we'll update

We'll rerun every benchmark once OpenRouter has a paid route, and add Artificial Analysis, Arena and the other sources as they publish. Until then, the MiniMax M3.1 Flash model page has every result as it lands.

Choosing or switching the model behind a product feature? Our custom evaluations test the models you are weighing on your own prompts and data, against your own definition of good. If AI writes your product content, the catalogue feed diagnostic measures how far you can trust it. Talk to us, or email hello@springprompt.com.

Quick answers