A cheaper token becomes useful when it helps finish the work. Compare the cost of an accepted result, including retries and repairs.
What is available now
MiniMax’s October 19 announcement launches M3.1 Flash through its API, M Plan and MiniMax Code. The company identifies the earlier anonymous Space Bunny service as a version of this model.
The announcement promises model weights later. It does not give a release date. Hosted access and downloadable weights therefore remain separate availability questions. Check the actual model selected in your application when moving from a preview.
Use the published rates as a starting estimate
MiniMax lists these rates per million tokens: $0.15 for input, $0.60 for output, and $0.015 for cached input. Those are the company’s published prices, not a measurement of a complete workflow.
For an illustrative request with one million uncached input tokens and 100,000 output tokens, the token cost is $0.21. That is $0.15 plus $0.06. It excludes tools, infrastructure, retries, and human review.
Record which input actually qualifies for caching. A cached-input rate does not price the parts of a prompt that changed, and a subscription allowance is different from a metered API bill.
Treat benchmark and speed claims as vendor results
MiniMax reports improvements in coding, computer use and other agent tasks, along with faster inference and lower task costs than its predecessor. We have not reproduced those results.
A tokens-per-second figure cannot establish how quickly an agent completes your task. The number of calls, permission requests, failed actions, and corrections can all change the result.
Define an acceptance test before switching
I would start with a small repository repair with clear expected behavior. Give each candidate the same issue, source revision, tool permissions, and relevant checks.
- Save the requested model and the available version information.
- Record all token categories and external tool charges.
- Count unsuccessful attempts as part of the workload.
- Review the patch for correctness and unrelated changes.
- Divide the total evaluation cost by the number of accepted repairs.
Keep human review time beside that calculation. A low-cost answer that needs substantial repair may be a poor default even when its token rate is attractive.
The updated agent cost-control guide covers budgets and routing. Our DeepSeek Flash guide separates native API prices from other providers’ terms.
Read the sources and plan your next step
What result would count as an accepted task in your first comparison?
Checked October 20, 2026. This is source-based reporting and analysis; proposed checks are not hands-on test results.
