MiniMax M3.1 Flash Review: A $0.15 Agent Model and the Cost Claims Checked
MiniMax says M3.1 Flash cuts agent costs by 70% at $0.15 input and $0.60 output. We check the price, the benchmarks and the weights promise.
Jump to 10 sections
MiniMax M3.1 Flash is an agent and coding model released October 19, 2026 at $0.15 per million input tokens and $0.60 per million output tokens. MiniMax says it cuts overall agent costs by 70% versus its predecessor. Weights are promised later, and the figures here are vendor-reported unless noted.
This review is for teams building coding agents or tool-using assistants who want a cheap, capable model and need to know what holds up. It covers price, the benchmark table, the cost claim and what remains unconfirmed. Model Drop reviews launches by separating what a vendor says from what independent data shows, and M3.1 Flash has both.
What Did MiniMax Actually Release?
MiniMax released M3.1 Flash on October 19 for agentic workflows, coding, computer use and cybersecurity. It is available through MiniMax's API, its M Plan and MiniMax Code. Pricing is $0.15 per million input tokens, $0.60 per million output tokens and $0.015 per million cached input tokens, according to Let's Data Science.
The model had an earlier life as a preview. An eesel guide says a preview appeared in MiniMax Code on September 27, before any official announcement, and carried no published per-token price at that stage. The October launch is the first time pricing is clear.
Technical details come from MiniMax's documentation as summarized by that guide, which describes a multimodal coding model with a 1 million token context window and tunable thinking depth. Input can be text, image or video, and thinking levels run from low to max, with max as the default.
How Much Does M3.1 Flash Cost to Run?
At list price, a job with 50 million input tokens and 10 million output tokens costs $13.50. That is $7.50 for input and $6.00 for output. If most of the input is cached, the input cost can drop sharply, since cached tokens cost one-tenth of the normal input rate.
The sibling model, MiniMax M3, is priced at $0.30 input and $1.20 output per million tokens, according to the eesel guide. M3.1 Flash is therefore half the price of M3 at list rates. Our overview of LLM API pricing in 2026 shows where that sits against the big labs.
The table below compares the price of the same job across the two MiniMax models.
| Model | Input rate | Output rate | Cost of sample job |
|---|---|---|---|
| MiniMax M3.1 Flash | $0.15 | $0.60 | $13.50 |
| MiniMax M3 | $0.30 | $1.20 | $27.00 |
Does the 70% Agent Cost Savings Claim Hold Up?
Partly, and it depends on how you count. MiniMax says the model reduces overall agent usage costs by 70% versus its predecessor. That is a vendor claim and the coverage we reviewed does not verify it.
Independent data points in the same direction but shows a smaller drop. The same Let's Data Science report cites Artificial Analysis as giving an average task cost of about $0.22, versus about $0.51 for MiniMax M3. That is a drop of roughly 57%, calculated from those two numbers.
The gap between 70% and 57% is not a contradiction. Vendors often measure their best case, while independent testers use a standard suite. A saving of more than half is still large. But it should shape how you plan, since a 57% cut and a 70% cut give different monthly bills.
What Do the Benchmarks Show?
MiniMax reports a set of agent and coding results. It lists 73.7% on DeepSWE v1.1, 61.7 on OSWorld V2, 29.6 on Agents' Last Exam and 82% on CyberGym. All four are MiniMax-reported, and benchmark setups differ between labs.
Independent results are thinner. The Let's Data Science report says Artificial Analysis gave the model an Intelligence Index score of 42. We have no direct comparison table for that score in the coverage we reviewed, so treat 42 as a single data point.
The eesel guide notes that MiniMax's architecture claims, such as sparse attention and faster decoding at long context, come from the M3 generation. Do not assume those numbers transfer to M3.1 Flash without a test.
Can You Trust Its Answers? The Reliability Flag
This is the weak spot. The coverage reports that 93% of the model's non-correct responses were incorrect answers, rather than partial answers or acknowledgments of uncertainty. That does not mean 93% of all answers are wrong. It means that when the model fails, it usually gives a confident wrong answer instead of saying it is unsure.
VentureBeat flagged a factual-reliability concern, according to the same report, and it recommends testing abstention behavior and factual accuracy before using the model in unsupervised workflows. That is good advice for any model, and especially for a cheap one that teams may run without review.
For agents, confident errors cost more than visible ones. An agent that says it is unsure can hand off to a human. An agent that is sure and wrong can write bad code or take a wrong action. Our guide to AI coding agents compared covers how teams set up review around such failures.
What About Weights and Open Access?
Weights are not yet released. MiniMax says they will come later, but the coverage gives no date. Some outlets have reported different dates or even called the model open-weight, which we could not confirm.
That matters for teams that want to self-host. Without weights, you must use MiniMax's API, its plans or a partner. Our comparison of open-weight and closed models explains how that choice affects cost, privacy and lock-in.
Until weights appear, treat M3.1 Flash as a closed API model. If a later release arrives, check the license terms before you plan around it.
How Fast Is It?
MiniMax says online output runs above 200 tokens per second, with a median above 170. That is a vendor figure. The eesel guide cites one day-one tester who measured roughly 90 to 110 tokens per second during the preview, which is a third-party measurement of an earlier build.
The two numbers may both be right. Speed depends on load, effort level and prompt length. A model set to max thinking will produce more thinking tokens, and total time per answer rises even if each token arrives quickly.
One practical detail from the documentation: thinking cannot be turned off. A request that disables thinking returns an error, so teams that want lower latency must lower the effort level instead.
Where Does M3.1 Flash Fit Among Cheap Agent Models?
It sits in the crowded low-price tier that every lab now serves. These models are meant for high-volume work where a premium model would be too expensive, such as triaging tickets, drafting code changes or running multi-step tool calls. The question is not whether they are as smart as the flagship. It is whether they are smart enough for the job at a price that makes the job worthwhile.
For coding agents, M3.1 Flash is aimed squarely at that role. MiniMax lists Claude Code, Cursor and Codex CLI as supported through its plans, which makes it easy to try inside tools many teams already use. That lowers the cost of a pilot, since you do not need to rebuild your workflow.
The trade-off is the usual one. A cheaper model may do well on routine tasks and fall behind on rare, hard ones. Many teams solve this by routing: easy requests go to the cheap model, and difficult ones escalate to a stronger one. M3.1 Flash is a candidate for the first tier, not necessarily the second.
Vendor lock-in is the other consideration. Without open weights, you depend on one company's API and pricing. A price that looks cheap today can change, so keep a fallback model tested and ready.
How Should You Test It Before Adopting?
Run a pilot on real tasks. Choose 50 to 100 prompts from your own workload, including a few where the correct answer is that the model should not know. Compare M3.1 Flash against your current model on accuracy, total tokens, time and cost per completed task.
Follow this short list during the pilot.
- Measure total tokens per task, not just the list price.
- Test at two or three effort levels to find your best trade-off.
- Include questions where the right answer is uncertainty.
- Log every wrong but confident answer and review them.
- Calculate cost per successful task, not cost per call.
That last measure is the most useful. A cheap model that needs three tries can cost more than a pricier one that gets it right once.
Verdict
MiniMax M3.1 Flash is an attractive price for agent work, and independent data supports a real cost drop versus M3, though not as large as MiniMax claims. The missing weights, the reliability flag on confident wrong answers and the lack of independent benchmark detail all argue for caution. Model Drop's editors would pilot it on low-risk tasks first, measure cost per successful task and keep human review on anything that matters.
