MiniMax M3.1 Flash costs $0.15 per million input tokens. And the Shanghai-based lab shipped it on October 19, 2026 through the MiniMax API and MiniMax Code, its developer-facing coding agent tool. MiniMax reports the model cuts usage costs by 70 percent versus the prior M3 generation while running nearly three times faster. It carries a 1-million-token context window, roughly 1,500 pages of text, with output priced at $0.60 per million tokens and cached input at $0.015 per million (Automater).
That is the summary.
Here is the part most coverage buries: the pricing is genuinely aggressive, the vendor benchmarks are strong. And one independent score in the mix tells you exactly where this model will break on you. Cheap tokens are only a win if the model finishes the task.
What Shipped, and What It Costs
| Token type | Price per 1 million tokens |
| — | — |
| Input | $0.15 |
| Output | $0.60 |
| Cached input | $0.015 |
The flat structure matters more than the headline number.
There is no tiered math, no per-domain surcharge, no reasoning-token surcharge visible in the pricing. You pay $0.15 in, $0.60 out, and a tenth of that for cache reads.
For comparison shopping, The Decoder pegs the cost of running one Artificial Analysis Intelligence Index evaluation task at about $0.22. Because the model uses far fewer output tokens than its rivals. That is the number I care about.
Per-token prices are marketing; cost per completed task is what shows up on your invoice.
The Benchmarks: Read Vendor Rows and Independent Rows Separately
MiniMax’s own numbers are loud. The company reports 73.7 percent on DeepSWE v1.1, the benchmark for resolving real GitHub issues end-to-end. It claims 76.2 percent on BankerToolBench, an agentic finance-tool benchmark, 61.7 on OSWorld V2 for Computer Use, 62 percent on Automation-Bench. And 91 percent on BabyVision for visual reasoning (Tech-Insider, Automater).
Then the same reporting hands you the number that should slow you down: 30 percent on ALE, the harder agentic exam, from the same vendor announcement.
That spread is the whole story. A model that scores 76.2 percent on a finance-tool suite and 30 percent on the general agentic test is a specialist, not a generalist. It will look brilliant inside the lane it was tuned for and quietly fail on long, ambiguous, multi-step chains where nobody defined the tools in advance. Small operators get burned by exactly this pattern: a model aces the demo task, then the first real workflow with eleven steps and two judgment calls eats the margin.
The independent rows are more encouraging.
Artificial Analysis testing, reported by The Decoder, scores MiniMax M3.1 Flash at 42 on the Intelligence Index, up 13 points from M3’s 29. It hits 48 percent on Humanity’s Last Exam, beating both DeepSeek and GLM.
And answers 56 percent of AA-Omniscience factual questions correctly against 43 percent for DeepSeek and 28 percent for GLM-5.3-Flash.
On CritPt, a physics-reasoning benchmark spanning 71 problems, it manages 22 percent.
Nobody should buy a budget agent for research physics, and that score says nobody will.
One trust note: coverage splits on DeepSWE, with one outlet reporting 73.7 percent and another rounding to 74. Same result, different rounding, and a reminder that launch-week numbers travel fast and get sloppy.
The $0.015 Cache Line Is the Real Pricing Story
Everyone quotes the $0.15 input price. The number that changes your architecture is the $0.015 cached input price.
Agentic work is repetitive by construction. Every step of a loop re-sends the system prompt, the tool definitions, the accumulated transcript. On a twenty-step agent run, most of what you pay for is the same context arriving twenty times. At full input price that is brutal. At a tenth of input price, long chains stop being the budget killer they are on most providers.
This is where MiniMax’s claimed 70 percent cost cut and the $0.22-per-task estimate start to feel coherent rather than promotional.
Fewer output tokens plus cheap cache hits compounds fast. If you build agents that loop, model your bill on cache behavior, not on the sticker input rate, given that the two diverge by an order of magnitude on real workloads.
What I’d Do With It as a Small Operator
First, do not switch your default model on launch week. Vendor benchmarks at release are the best-case row of the spreadsheet. Route a fixed batch of your real tasks through the API side by side with your current model and score the outputs yourself before anything touches production.
Second, match the model to the failure cost. The 30 percent ALE score is a warning label for unsupervised long chains. Use MiniMax M3.1 Flash where a retry costs pennies: draft generation, summarization, first-pass code review, classification at volume. Keep an expensive-to-verify judgment call on whatever you already trust, as a cheap confident failure is still a failure you have to catch.
Third, treat the 1-million-token context window as a claim, not a feature. The figure of roughly 1,500 pages is MiniMax’s own framing. Before you point it at a codebase or a document archive, test retrieval accuracy on documents you already know cold. Context windows that technically accept a million tokens do not always usefully attend across all of them. And a missed detail in a long document is worse than a truncated one since it fails silently.
The bigger picture for small teams: the price floor for competent agentic models keeps dropping. And this launch is another leg down. Your use is not loyalty to any lab. It is knowing your own cost per completed task, on your own work, better than the vendor knows its benchmarks. The operators who win the next year of AI pricing are the ones running that math monthly, not the ones chasing every launch.
Pull your last API invoice this week and compute cost per finished task, not per token.
Then run one real workflow through MiniMax M3.1 Flash before you decide what it replaces. Cheap and fast only counts when the task actually completes.
