MiniMax M3.1 Flash tested: cheap per token, but not per coding job
We gave MiniMax's new M3.1 Flash our secret coding test. It passed 175 of 186 checks, but wrote so much per answer that a job cost more than with Claude Haiku 5.5.
MiniMax released M3.1 Flash on 19 October 2026. It is a new AI model for coding, and it is very cheap per token. We gave it our secret coding test, together with its bigger brother MiniMax M3. M3.1 Flash passed 175 of 186 checks. That is good, but it wrote so much per answer that one job cost more than with Claude Haiku 5.5, which passed every check.
What is MiniMax M3.1 Flash?
MiniMax is an AI company that makes models for text, images, audio, video and music. On 19 October it announced M3.1 Flash, a cheaper "Flash" model next to its main model, M3.
MiniMax says M3.1 Flash is "nearly three times faster" than before. It says the typical speed it measured is more than 170 tokens per second. The model is sold through an API, the way apps and coding tools talk to a model. On MiniMax's price page, it costs $0.15 per million input tokens and $0.60 per million output tokens. That is half the price of M3, which costs $0.30 and $1.20.
MiniMax says "model weights will be released later". Weights are the model files, so you could then download it and run it yourself. MiniMax also reports 73.7% on a coding test called DeepSWE. And it says M3.1 Flash ranks first among Chinese models on two tests for AI agents. We did not check those numbers; we ran our own test.
Our test
We used our secret coding test, the same one we give every model. It has three jobs: fix the money maths in a small program, write a bill splitter, and write a CSV reader, which turns a text file with commas into a table. Each job came in one message, and the model had to answer with the full code. Hidden tests then checked it: 186 checks over three tries per job. Each answer could take at most 15 minutes. You can read more in how we test AI models.
- M3.1 Flash: we used the free version on OpenRouter, a service that passes requests on to AI companies. MiniMax itself ran every request.
- M3: the normal paid version on OpenRouter. Other companies ran it for MiniMax: Together and CoreWeave.
We compare both with earlier runs on the same test, such as Claude Haiku 5.5, GPT-6 Luna, GLM-5.3 Flash and DeepSeek V4 Pro.
How good was the code?
M3.1 Flash passed 175 of 186 checks. 7 of its 9 jobs were perfect. M3 passed 178, with 5 perfect jobs.
The CSV reader went well: Flash got all 17 checks right in all three tries. The money maths was perfect twice; once it missed 2 checks on unusual number formats. The bill splitter was the weak spot. One try got only 8 of 17 checks, with mistakes in tips, leftover cents and order. The other two tries were perfect.
That is a good score for a cheap model. But it is not the best on our test. Claude Haiku 5.5 and Gemini 3.8 Flash passed all 186 checks. GPT-6 Luna and GLM-5.3 Flash passed 184.
Cheap per token, not per job
M3.1 Flash ran for free in our test. To compare it fairly, we worked out what the same tokens would cost at MiniMax's own price. That came to about 1.5 cents per coding job, or $0.14 for all nine.
That sounds tiny, and it is. But four models that got more checks right were cheaper per job:
- GPT-6 Luna: about 0.12 cents per job (184 checks),
- Claude Haiku 5.5: about 0.19 cents (186 checks),
- GLM-5.3 Flash: about 0.69 cents (184 checks),
- DeepSeek V4 Pro: about 1.4 cents (180 checks).
MiniMax M3 cost more: OpenRouter billed us about 2.3 cents per job. At MiniMax's own list prices for both, M3.1 Flash was about 37% cheaper per job than M3. MiniMax says it is about 70% cheaper for agent tasks; we did not test agent tasks. See our AI API price comparison for more models.
Why: it writes a lot
The reason is length. Before it writes code, M3.1 Flash "thinks" in hidden text. In our test, it wrote about 20,000 tokens per answer. Claude Haiku 5.5 wrote about 3,000 and GPT-6 Luna about 2,000 for the same jobs.
It is not the longest writer: M3 wrote about 18,000 per answer and Gemini 3.8 Flash about 27,000. In total, Flash wrote 230,000 tokens over nine jobs and M3 183,000. Flash is cheaper than M3 because its price per token is half, not because it writes less.
Some answers were very long. One bill splitter took 47,000 tokens, and one CSV reader 55,000. You pay for every one of those tokens, even the hidden ones. So a low price per token helps most when a model does not think too long.
How fast was it?
On the free version, a typical job took about 215 seconds. The slowest took 716 seconds. If you divide the tokens by the whole time per answer, waiting included, that is about 67 tokens per second.
MiniMax's number of more than 170 tokens per second is measured in a different way, on its own service. So the two numbers do not compare directly. The paid API might be faster than the free version; we did not test it.
The paid M3 took about 116 seconds per job in our test. Claude Haiku 5.5 took about 15 seconds: it wrote much less, and wrote it faster.
Where your data goes
MiniMax's API privacy policy, in effect since 30 March 2026, names Nanonoble Pte. Ltd., a company in Singapore, as the one that handles your data. It says your data is "stored in the data center located in the United States".
The policy says MiniMax processes what you send and what the model answers, to provide its AI services. It also says MiniMax may de-identify data, which means removing anything that shows who it came from. It may then "mine, analyze and utilize" that data commercially. The policy says MiniMax does not use personal data "for training to profile or target consumers". It does not clearly say whether your prompts in general are used to train models. If you send private code, keep that in mind.
If you use M3.1 Flash through OpenRouter or Vercel, your text also passes through them on its way to MiniMax.
Where you can try it for free
- Vercel AI Gateway: free until 22 October, but only when MiniMax itself runs the request. Vercel advises you to set MiniMax as the first provider, the company that runs the model, to make sure.
- OpenRouter: a free version, run by MiniMax. This is the one we used. Free versions can be slower or have limits.
- MiniMax's own API: $0.15 in and $0.60 out per million tokens. Its price page lists no free offer.
Free offers can end at any time. If you build something on M3.1 Flash, plan for the normal price.
Should you use it?
- You want the cheapest code that passes? On our test, Claude Haiku 5.5 and GPT-6 Luna did better for less money per job.
- You want a model to run yourself? M3.1 Flash may be interesting once its model files are out. MiniMax says they come later. To run a model at home, see our LM Studio vs Ollama test.
- You need answers fast? The free version took minutes per job.
- You just want to try it? It is free right now on Vercel and OpenRouter, and the results were good.
M3.1 Flash is a good, cheap model. It is just not as cheap per job as its price per token suggests.
Questions people ask
Is MiniMax AI free or paid?
Its API is paid: M3.1 Flash costs $0.15 per million input tokens and $0.60 per million output tokens at MiniMax. Right now you can try M3.1 Flash for free: on Vercel's AI Gateway until 22 October, and as a free variant on OpenRouter, which we used. We did not check the prices of MiniMax's own apps.
How much does MiniMax M3 cost?
On MiniMax's price page, M3 costs $0.30 per million input tokens and $1.20 per million output tokens, for prompts up to 512,000 tokens. MiniMax marks this as "Permanent 50% off". In our test, M3 cost about 2.3 cents per coding job, or $0.21 for all nine jobs.
Is MiniMax M3 really good?
It did well, but not perfectly. On our secret coding test it passed 178 of 186 checks, and 5 of 9 jobs were perfect. M3.1 Flash passed 175. Claude Haiku 5.5 passed all 186, faster and for less money per job.
Is MiniMax M3 completely free?
No. M3 is a paid model on MiniMax's API and on OpenRouter. The new M3.1 Flash is the one you can try for free right now, on Vercel's AI Gateway until 22 October and as a free variant on OpenRouter.
Why is MiniMax so cheap?
MiniMax says M3.1 Flash uses "more aggressive sparsity" to run more efficiently; it does not explain more. But a low price per token does not always make a job cheap. In our test, M3.1 Flash wrote about 20,000 tokens per answer, so one job cost about 1.5 cents at its list price.
Is MiniMax a Chinese model?
MiniMax's own announcement compares M3.1 Flash with other Chinese models and says it ranks first among them on two agent tests. The pages we read do not say where the company is based. Its API privacy policy names Nanonoble Pte. Ltd., a company in Singapore, as the one that handles your data. It says the data is stored in a data center in the United States.
What is MiniMax used for?
MiniMax makes AI models for text, audio, images, video and music. According to its investor page, they power MiniMax Agent, Hailuo AI, MiniMax Audio, Talkie and an API for developers. We only tested its text models on coding.
How we checked
On 21 October 2026 we read MiniMax's announcement of 19 October, its price page, its API privacy policy and its investor page. We also read Vercel's changelog and OpenRouter's model list.
Then we ran our secret coding test v2 through OpenRouter. It has three jobs, three tries each, and 186 hidden checks. Each job was one request, with a limit of 15 minutes per answer. M3.1 Flash ran on the free version (minimax/minimax-m3.1-flash:free), served by MiniMax. M3 ran on the paid version (minimax/minimax-m3), served by Together and CoreWeave. Time and token counts come from OpenRouter's replies.
For M3, OpenRouter billed $0.21 in total, a bit below MiniMax's list price. M3.1 Flash cost us nothing, so we worked out its cost from its tokens at MiniMax's list price. For the Flash and M3 comparison, we used list prices for both. The other models come from our earlier runs of the same test. The charts are our own, and none of the images was made by AI. Three tries per job is a small test, and free routes can change, so your results may differ.