MiniMax M3.1 Flash went on sale on 19 October, and the Shanghai company is pitching it on two numbers: a list price of $0.15 per million input tokens and output speeds above 200 tokens per second. The model is aimed at AI agents, software development, computer use and security work, and it is available now through the MiniMax API, the M Plan subscription and the MiniMax Code desktop app.
For developers in Africa and elsewhere who pay for every token their agents burn, the price is the headline. Here is what MiniMax has published, what it has left out, and where the model sits among this month's cheap, fast releases.
MiniMax M3.1 Flash pricing
MiniMax lists a flat rate for M3.1 Flash:
- Input: $0.15 per million tokens
- Output: $0.60 per million tokens
- Cached input: $0.015 per million tokens
The cached-input rate is a tenth of the normal input price. That matters for agents, which tend to resend the same long system prompt and tool definitions on every step. MiniMax says the cost of running agentic tasks is 70% lower than with the model's predecessor, though it does not publish the workload behind that figure.
Speed claims
The company says output "reaches 200+ tokens per second in everyday use", with median speeds above 170 tokens per second, and that online inference is close to three times faster than before. It credits "more aggressive sparsity" for the gain, which suggests a mixture-of-experts design that activates a smaller share of its parameters per token.
MiniMax has not disclosed the parameter count, the architecture in detail or the context window. Those gaps make independent comparison harder, and buyers should test latency on their own prompts before committing.
An anonymous trial run
One detail in the launch post stands out. MiniMax says an early version of M3.1 Flash ran anonymously as "Space Bunny" and processed more than 200 trillion tokens before the official release. Stealth testing of this kind has become common among model makers, who use public leaderboards and routing services to gather real usage before putting their name on a model.
Benchmark results, as reported by MiniMax
All of these scores come from the company itself and have not been independently reproduced:
- BankerToolBench: 76.2
- DeepSWE v1.1: 73.7%
- OSWorld V2: 61.7, which MiniMax says is first among Chinese models
- Agents' Last Exam: 29.6, also first among Chinese models by MiniMax's account
- CyberGym: 82%
MiniMax also says the model ranks "among the top performers" on AutomationBench-AA without giving a number. The mix of tests says a lot about the intended use. OSWorld measures how well a model operates a desktop computer, DeepSWE covers multi-step software engineering, and CyberGym tests security tasks.
What the model is built for
The launch post lists three groups of uses:
- Enterprise workflows that run over many steps in sales, marketing, operations, customer service, finance and HR.
- Long software projects, where an agent writes, tests and fixes code over an extended session.
- Security work, including code auditing, vulnerability patching, penetration testing, malware analysis and reverse engineering.
The security list is broad. Tools that can patch a vulnerability can usually help find one too, and companies deploying the model for that work should keep their own logging and access controls in place.
"Flash represents our ongoing exploration of the balance between efficiency and capability," the company wrote.
How it compares with this month's other fast models
October has been crowded at the cheap end of the market. On the same day MiniMax shipped, Inception released Mercury 3, a diffusion model it says passes 1,000 tokens per second. We covered Mercury 3's launch and its $0.20 per million pricing earlier today. DeepSeek also updated its V4.1 Flash line on 17 October.
The practical difference is focus. Mercury 3 sells raw speed, while MiniMax is selling an agent model that is cheap enough to leave running. Its $0.60 output price is the number to watch for agents that write long answers or large amounts of code.
Availability and open weights
M3.1 Flash can be used today through three routes:
- the MiniMax API, documented on the company's developer platform
- the M Plan, MiniMax's subscription for regular users
- MiniMax Code, its desktop coding app
MiniMax says it will release the model weights later but has not given a date or a licence. It also says a larger "Pro" model is coming. Until the weights appear, teams that need to run models on their own servers will have to wait.
MiniMax's launch post does not say anything about regional availability, so developers in African markets should check sign-up and payment options on the platform directly rather than assume access.
Should you switch?
For teams already paying for a mid-range model to run agents, M3.1 Flash is worth a test, mainly for its cached-input price. The case against moving fast is the missing detail: no context length, no architecture disclosure, no independent benchmarks and no licence yet for the promised weights. A week of side-by-side trials on real tasks will tell you more than any of the published scores.
