MiniMax M3.1 Flash Explained: Speed, Pricing & AI Agent Performance

AI developers often find that a fast model is not the same as a cheap or reliable agent. A model can generate tokens quickly, yet an agent that retries tool calls, rereads long files, and loops through a task can cost more and take longer than the headline numbers suggest. MiniMax M3.1 Flash, a preview model from the Chinese AI company MiniMax, is positioned for fast, low-cost coding and agent work. As of October 20, 2026, though, public information about it is thinner than its name implies. This article separates what MiniMax has documented, what independent observers have reported, and what remains unverified, so you can decide whether it deserves a test.

MiniMax M3.1 Flash at a Glance

MiniMax M3.1 Flash, published under the API model ID MiniMax-M3.1-Flash-Preview, is a preview of a coding-oriented model in MiniMax’s M3 series. MiniMax’s developer documentation describes it with a 1,000,000-token context window and always-on reasoning. Third-party trackers list its release as September 27, 2026, and a Chinese technology review reported that it reached MiniMax Code and the Token Plan the following day. I could not confirm the October 19 rollout date that some coverage cites.

A model generates output from the input it receives. An AI agent wraps a model with tools, instructions, an execution loop that carries out actions, and often memory or verification steps. Benchmarks measure specific combinations of these pieces, so a model’s name alone says little about how it will perform inside your agent.

How Fast Is MiniMax M3.1 Flash?

Speed in AI has several meanings, and vendors often quote only one. Output throughput, measured in tokens per second (TPS), describes how quickly the model writes after it starts. Time to first token describes how long you wait before any output appears. Total response latency covers the whole reply, and task completion time covers everything an agent does, including tool calls, file reads, retries, and checks. A model with high throughput can still finish a task slowly if it makes many calls or generates long reasoning traces.

MiniMax documents the effort setting as a speed and depth trade-off. The default max level produces the most reasoning tokens, and typically the longest waits. Lower levels reduce that cost but may reduce answer quality on hard problems.

Independent observations are limited. A Chinese technology reviewer who tested the model in MiniMax Code reported output that stayed near 100 tokens per second regardless of how long the model thought. The same reviewer wrote that fast output did not mean the test projects finished quickly. Another analysis cited a decode range of 90 to 110 tokens per second, but it explicitly described those as numbers from the M3 generation, not measurements of M3.1 Flash. I found no MiniMax speed figure for this preview. Treat any specific TPS number for this model as unverified until MiniMax or a reproducible test publishes one.

Speed also varies with prompt length, output length, concurrency, cache hits, and provider load. Measure it on your own prompts rather than relying on a single headline number.

MiniMax M3.1 Flash Pricing: What Does It Cost?

The most important pricing fact is that I found no published per-token rate for M3.1 Flash Preview. Access runs through subscription plans, and subscriptions do not establish a public per-token price. AI Agent Store notes that a subscription bill can also include tools, retries, and charged thinking tokens, so subscription access alone does not tell you what an agent workload will cost.

For a reference point, MiniMax’s pay-as-you-go pricing for the M3 model, as recorded on Activepieces’ October 15, 2026 review of MiniMax pricing, is as follows for requests up to 512,000 tokens:

Above 512,000 tokens, the same review lists $0.60 input, $2.40 output, and $0.12 cache read. These figures belong to M3, not M3.1 Flash Preview. Do not apply them to the preview without confirming current pricing on the MiniMax platform.

Here is a transparent calculation using the M3 rates above, with stated assumptions. Suppose a task sends 100,000 input tokens and receives 20,000 output tokens, with no caching:

• Input: 100,000 × $0.30 / 1,000,000 = $0.030
• Output: 20,000 × $1.20 / 1,000,000 = $0.024
• Total: $0.054

If 80,000 of the input tokens are cache hits, the same task changes as follows:

• Cached input: 80,000 × $0.06 / 1,000,000 = $0.0048
• Uncached input: 20,000 × $0.30 / 1,000,000 = $0.0060
• Output: $0.024
• Total: $0.0348

Caching matters most in agent loops, where the same instructions, tool definitions, and file context are resent on every turn. Total cost per task also depends on how many calls the agent makes and how much reasoning it performs. For that reason, track cost per successful task rather than cost per token.

MiniMax’s documentation includes a related claim. Its Codex integration guide reports that in a ProgramBench comparison, a 512K context window produced similar average performance to a 1M window at an estimated cost about 20 percent lower. That is a vendor-documented comparison on one benchmark, so use it as a prompt to test context sizes on your own workload.

Benchmarks and AI Agent Performance

Benchmark claims need careful handling. I could not find a MiniMax benchmark table for M3.1 Flash Preview in the documentation and sources I checked. The figures in some circulating coverage, including specific scores on coding, computer-use, and finance benchmarks, could not be traced to a primary source, so this article does not repeat them.

What can be checked is the evidence for the M3 model, which predates this preview. Vals AI’s M3 page lists a Vals Index accuracy of 58.94 percent, with a plus or minus 1.82 margin, a latency of 1,477.67 seconds, and a cost of $1.50 per test. The BenchLM mirror of Vals Index v2 places M3 at 42.72 percent. These figures cannot be merged or compared directly, most likely because they come from different benchmark versions, and I could not confirm the version behind each. Check the Vals AI site directly before citing either number.

Coding and software engineering

Coding is the best-documented use case. MiniMax’s integration guides show the model working inside Claude Code, Codex, Cursor, and OpenClaw. The Codex guide points to the OpenAI-compatible base URL https://api.minimax.io/v1 and a model name, and it states that the no-reasoning effort is unsupported. The Claude Code guide uses the Anthropic-compatible endpoint https://api.minimax.io/anthropic and confirms that thinking cannot be turned off. These details matter in practice, because a configuration that assumes you can disable reasoning will fail.

A reviewer who tested the model in MiniMax Code reported that it handled five project-style tasks, including SVG animation and video-to-manual conversion, with consistent output. The same reviewer found that completion was not especially quick. This is one author’s testing, not a reproducible benchmark, but it reflects the trade-off readers should expect: steady output does not guarantee fast task completion.

Computer use, data analysis, and business automation

I did not find independent, published results for M3.1 Flash Preview on computer-use, data-analysis, or business-automation benchmarks. An evaluation service lists private task runs at each reasoning level, but I could not access the results. Until comparable results are published, the sensible approach is to build a small test set from your own workflows, including the failure cases you care about most.

Cybersecurity

I did not find verified cybersecurity benchmark results for this model, so this article makes no claims about its security performance. Any security testing should be restricted to authorized systems and reviewed by a human.

Real-World Use Cases: Where Might M3.1 Flash Fit?

The model’s documented strengths point toward routine coding work. For bug fixes, small features, and code review inside an existing repository, the main benefit is the ability to run the model in tools developers already use. The condition for success is the test suite: an agent that can run the tests and read the results is far more useful than one that only edits files. Keep reasoning effort at a level that matches the task, because max adds latency on simple edits.

For business automation across connected software, the model may suit multi-step workflows that call several APIs. The success condition is permission design. Give the agent the narrowest credentials that work, log every tool call, and require approval for actions that write to production systems.

For computer-use agents that operate browsers or desktop applications, the appeal is the model’s image and video input, which the documentation lists. The condition is isolation. Run these agents in sandboxed environments with restricted access, because a misread screen can lead to a harmful action.

For data analysis and reporting, the model can draft queries, summarize tables, and explain results. The condition is validation. Check every number against the source data before it reaches a stakeholder, because language models can produce confident but incorrect calculations.

In every case, the model alone does not provide a complete solution. Agent frameworks, tools, permissions, and verification steps determine whether the system succeeds.

MiniMax M3.1 Flash vs. MiniMax M3

The comparison shows that these models serve different needs at this point. M3 has a published price, open weights, and a longer public record. M3.1 Flash Preview is newer, always reasons, and is available only through subscription access. Because the preview has no published per-token price and no comparable benchmarks, a cost-based migration decision cannot be made from public data alone. Test both models on representative tasks, measure cost per successful task and time to completion, and migrate only if the results support it.

Limitations and Best Practices

The biggest limitation is the gap between what is documented and what has been independently measured. MiniMax’s documentation covers context size, endpoints, and reasoning controls, but I found no published speed or per-token price for the preview. Independent evidence is limited to a small number of reviews and private evaluations.

Latency and completion time are variable. The reviewer who tested the model found that fast output did not produce fast task completion, and this pattern is common to reasoning models in agent loops. Errors can also accumulate over long workflows, because a wrong early step affects every later step.

Costs can rise unpredictably. Repeated tool calls, retries, large contexts, and always-on reasoning all add tokens. Subscription pricing can hide these costs until a billing period ends.

Data handling and access need attention as well. Any agent that reads files, sends messages, or changes systems needs explicit permissions, audit logs, and human review for consequential actions.

Model status is also changing. Because this is a preview, endpoints, limits, and access rules may change. Recheck the MiniMax documentation before you build on it.

Practical steps help reduce these risks:

1. Benchmark on your own tasks, using real repositories, documents, or workflows.
2. Track cost per successful task, not cost per token.
3. Log every tool call and its result.
4. Set spending limits on API keys and subscriptions.
5. Validate outputs against source data before acting on them.
6. Test failure recovery by deliberately breaking tool calls and checking whether the agent stops safely.

Final Verdict: Is MiniMax M3.1 Flash Worth Trying?

MiniMax M3.1 Flash has a clear technical profile: a 1,000,000-token context window, multimodal input, always-on reasoning with adjustable effort, and documented integrations with major coding tools. What it lacks, as of October 20, 2026, is a published per-token price, an official speed measurement, and independent benchmark results that can be checked and compared.

Developers who already use MiniMax’s subscription plans or coding tools can run a short, controlled trial on real coding tasks at low cost. Teams planning production automation should be more cautious. They should wait for published pricing and reproducible evaluations, or run their own tests with strict spending and permission limits before depending on the model. The right next step is a small trial, measured against the workload you actually run.

Frequently Asked Questions

What is MiniMax M3.1 Flash?

MiniMax M3.1 Flash, API model ID MiniMax-M3.1-Flash-Preview, is a preview coding-oriented model from MiniMax. It has a 1,000,000-token context window and always-on reasoning.

How fast is MiniMax M3.1 Flash?

MiniMax has not published a measured speed for this preview. One independent reviewer reported output near 100 tokens per second in MiniMax Code, but task completion was not especially fast.

How much does MiniMax M3.1 Flash cost?

I found no published per-token price for the preview. Access is through M Plan subscriptions. MiniMax’s pay-as-you-go rates for the older M3 model are $0.30 per million input tokens and $1.20 per million output tokens.

Is MiniMax M3.1 Flash good for AI agents?

Its documented features suit agent work, including the large context and tool integrations. Public benchmark evidence for agent tasks is limited, so testing on your own workflows is necessary.

What are the MiniMax M3.1 Flash benchmark results?

I could not verify official or independent benchmark results for the preview. The published Vals AI and BenchLM figures cover the M3 model and use different benchmark versions.

How does MiniMax M3.1 Flash compare with MiniMax M3?

M3 has published pricing, open weights, and a longer public record. M3.1 Flash Preview is newer, always reasons, and is available only through subscription access.

Can developers access MiniMax M3.1 Flash through an API?

MiniMax’s documentation describes Anthropic-compatible and OpenAI-compatible endpoints, but the preview requires an M Plan Subscription Key. That key is not interchangeable with a pay-as-you-go API key.

Are MiniMax M3.1 Flash model weights available?

Not as of the sources I checked. Third-party listings report no released weights for the preview, while M3 is described as open-weight.