A cheap AI answer is lovely. A cheap AI answer that needs three corrections, another tool call and someone to fix the spreadsheet is a small administrative adventure.
MiniMax’s 19 October launch of M3.1 Flash gives that familiar problem a fresh price tag. The company claims almost three times faster inference and 70% lower overall usage cost than its predecessor. It says the model is available through its API, M Plan and MiniMax Code, with downloadable weights to follow.
Our view: the rates are worth investigating. The more useful purchasing question is how much it costs to finish your particular job, at an acceptable standard, without a person quietly doing the difficult part.
What this article is: an analysis of primary documentation checked on 20 October 2026. We have not run MiniMax M3.1 Flash, measured its speed or independently reproduced its benchmark results. The calculations below are illustrative.
A rate card worth reading properly
MiniMax M3.1 Flash API prices
Published pay-as-you-go rates, USD per million tokens.
- Uncached input
- $0.15
- Output
- $0.60
- Cached input reads
- $0.015
Source: MiniMax’s pricing documentation. Subscription allowances and third-party routes are separate offers. Rates may change; these are not a promise about your final invoice.
The same rate card lists standard MiniMax M3 at $0.30 input and $1.20 output per million tokens for prompts up to 512K input tokens. On those two uncached categories, Flash’s listed rates are half as much. M3 also has a higher input-length band and a separately priced priority tier, so comparisons need to name the configuration.
That does not disprove the launch’s 70% usage-cost claim. It means a per-token discount and the cost of completing a task answer different questions. Token consumption, cache use and how much work gets done can change the result. The headline percentage alone does not establish a saving for your workload.
The token meter is only one meter
For a coding assistant, the deliverable is a change that meets the brief and passes meaningful checks. For document work, it might be a report with correct figures and usable references. Producing a plausible-looking file is the halfway point at best.
Consider a purely invented batch of 100 small jobs. Each first attempt consumes 10,000 uncached input tokens and 2,000 billed output tokens. At the Flash rates above, that is $0.0015 plus $0.0012: $0.0027 per attempt, or $0.27 for the batch. This excludes tools, infrastructure, taxes and human time. It assumes all billed output is captured in the stated token count.
Illustrative arithmetic · not model performance
| Accepted jobs from 100 attempts | Token cost per accepted job |
|---|---|
| 100 | $0.0027 |
| 50 | $0.0054 |
| 25 | $0.0108 |
Each row divides the same $0.27 by a different hypothetical acceptance count. No retries are included. These are not observed success rates for MiniMax or any competing model.
The point is not that Flash will fail a particular share of jobs. We do not know that. The point is that counting successful work changes the denominator. Retries add tokens; checking and repairing results adds time. At these small token costs, a few minutes of avoidable correction can matter more than the model bill.
Check the thinking setting before the stopwatch
The current invocation guide lists MiniMax-M3.1-Flash as the API model identifier, distinct from older Preview references. Its documented context capacity is one million tokens, which is a capacity limit rather than evidence that it reliably uses every part of a long input.
The guide also documents thinking effort from low through max, with max as the default. Thinking cannot simply be disabled; the documented way to reduce its latency and token use is to lower the effort. A default-versus-default comparison can therefore conceal different amounts of reasoning work.
If you compare models, keep the task and acceptance checks fixed, record the actual settings and compare the trade-offs. Try a lower effort on a bounded task before assuming the most elaborate setting is the most economical. It may save time, or it may need more corrections. Either outcome belongs in the record.
Benchmarks are useful leads, not your acceptance test
MiniMax’s launch presents coding, computer-use and automation benchmark results. Those are the company’s reported findings, not measurements made by AI The Thing. We are not turning that scorecard into an independent ranking.
There is also a practical gap between an API model and an agent product. Tools, prompts, permissions, retry rules and validation all affect what gets delivered. Changing the engine does not tell you whether the rest of the vehicle is roadworthy.
The documentation’s description of Flash as open-weight also needs a timing qualification: the dated launch says weights will be released later. API access is not evidence that the downloadable artifact, its licence and a workable local deployment are available today. We would verify those separately before suggesting a self-hosted setup.
What would a useful test look like?
We would start with harmless synthetic tasks: repair a small function against predefined tests, extract facts from a supplied document, or reconcile a table with deliberately planted discrepancies. Define success before running the model. Keep the inputs, tool access and stopping rules comparable, then repeat the tasks rather than showcasing the prettiest attempt.
- Accepted work: did the result meet the brief, including edge cases and factual checks?
- Full token usage: record the provider’s billed input, cache and output categories across every attempt.
- Time to acceptance: include tool execution, retries and necessary human correction, not just output speed.
- Failure behaviour: retain timeouts, incorrect outputs and tasks the model could not finish.
- Execution conditions: record the model identifier, effort, date and route, with bounded tool permissions and a spending limit.
This is a proposed test plan. No such comparison has been performed for this article. It would produce a useful case study, not a universal league table.
For now, MiniMax M3.1 Flash earns a closer look on price. A sensible trial would ask it to complete a small, representative piece of work and keep an honest account of the cleanup. The bargain is the finished job. Everything before that is a quote.
Related reading: our analysis of agents in production, inference-cost overview and look at evidence behind AI answers.
Sources and editorial method
This article draws on MiniMax’s 19 October release announcement, API pricing, model invocation guide and model catalogue, checked on 20 October 2026. Vendor claims are attributed; the cost example is our arithmetic using explicitly hypothetical workloads. Availability for a particular account or region has not been tested.
AI tools assist our research and publishing workflow, with strict controls around human input, review and editorial approval. No original model-performance results are presented.