MiniMax M3.1 Flash: a real price cut, a rented kind of fast
MiniMax's M3.1 Flash ships at $0.15 per million input tokens with strong agent scores, but the weights are not out and every benchmark is self-scored.
Vlad Makarovreviewed and published
MiniMax spent most of October serving a model it would not name. Space Bunny, as the anonymous endpoint was called, processed more than 200 trillion tokens before anyone knew whose it was, and on October 19 the company said so in a sentence that reads like an afterthought: Space Bunny was an early version of MiniMax M3.1 Flash. The trillions are the best number in the release, because they describe a model already in production, already billed, and already judged by people who had no idea what they were talking to.
Everything else needs a closer look. The weights are not out, every benchmark on the page was scored by the company that built the model, and the speed claim describes MiniMax's own serving stack rather than a property of files you can download. That does not make the release uninteresting. It makes it a pricing event with a capability story attached, and the two deserve to be read separately.
What shipped, and where you can actually reach it
The model is live through the MiniMax API, the M Plan subscription and MiniMax Code, the company's desktop agent app. It takes text, image and video in, returns text, and runs a 1M-token context window, per the product page. Reasoning is always on and cannot be switched off, with five effort levels from Low to Max, Max the recommended default. That is a change from M3, where reasoning was optional, and it matters when porting prompts: answers now arrive with a long thinking trace attached.
Pricing is flat per million tokens, with no long-context surcharge printed on the launch page.
| Item | Price per 1M tokens |
|---|---|
| Input | $0.15 |
| Output | $0.60 |
| Cached input | $0.015 |
Cached reads cost a tenth of the input rate, which decides whether an agent loop is affordable: repeated system prompts and file context get cheap, fresh output does not. A free endpoint also sits on OpenRouter, listed at zero dollars in and out. Self-hosting is the gap. MiniMax says weights will arrive later, and that a larger Pro model will follow, with no dates on either.
The benchmark sheet is signed by the vendor
The numbers are good, and they are all MiniMax's.
| Benchmark | Score | What it measures |
|---|---|---|
| AutomationBench-AA | 62.2 | Agentic SaaS workflows |
| BankerToolBench | 76.2 | Multi-step finance and tool chains |
| DeepSWE v1.1 | 73.7% | Repository-level software tasks |
| OSWorld V2 | 61.7 | Computer use across applications |
| Agents' Last Exam | 29.6 | Long professional tasks |
| CyberGym | 82% | Vulnerability discovery |
The claim travelling with the table is that M3.1 Flash ranks first among Chinese models on both OSWorld V2 and Agents' Last Exam. That may be true; it is also a claim about a subset, chosen by the seller, on a harness the seller controls, and the product page's FAQ concedes that these results describe the cited evaluations rather than guaranteed outcomes.
One independent read exists. Artificial Analysis scores the model at 43 on its Intelligence Index, fourth of 173 in its class against a median of 13, at $0.22 per index task. Fourth of 173 is a serious result for a model priced in the cents, and it is the only score here MiniMax did not produce, which is why it belongs above the vendor rows.
Three times faster, on whose machines
MiniMax reports output speeds above 200 tokens per second, with median speeds consistently above 170. The phrase carrying that claim is online inference, and it is doing real work: a serving stack can be fast for reasons that never reach a customer, including batch sizes tuned to a benchmark, a prompt mix heavy on prefill, and hardware the vendor chooses. Three times faster than its predecessor is plausible given the sparsity work in the post, and unverifiable from outside until the weights exist.
The independent measurement is 104.7 output tokens per second, below Artificial Analysis's median of 114 across 151 models and described as slower than average. Both are honestly obtained and measure different things: a best-case serving configuration against a shared endpoint under a standard evaluation. Read 200 tokens per second as a deployment, not as a property of the model.
428 billion parameters, 23 billion of them awake
The launch post never prints a parameter count, a small tell in a release this detailed. The model it descends from does: the M3 card lists about 428B total parameters and about 23B activated, across 128 experts with four live per token. Secondary write-ups report the Flash at the same totals, and the suffix changes the second number, not the first.
Weight memory scales with 428B: roughly 856 GB of tensors at bf16, 428 GB at eight bits, 214 GB at four bits per weight, all before any KV cache for a million-token context. Compute per token scales with the 23B that fire, which is how serving gets cheap and how MiniMax reaches 200 tokens per second on its own hardware. A Flash badge here is a claim about the bill, not about your laptop. Hacker News put it more bluntly, in a comment that also does the arithmetic: "23b active parameters isn't flash in my book. At 428B it can still fit on 2x 128GB machines at 4bpw, which is nice."
The cost claim that survives contact
The headline is 70% lower overall usage cost than the predecessor on agentic tasks, and the product page is careful about what that means: a task-cost comparison, not an API unit-price comparison. A task costs whatever the model charges multiplied by the tokens it burns, and the sparsity increase is what buys the speed.
The per-token bargain is also complicated by verbosity. Artificial Analysis measured 180 million output tokens across its index against a median of 100 million, and calls the model very verbose. Input at $0.15 against a $0.24 median and output at $0.60 against $0.87 still leaves room, but an 80% token overhead eats most of a 40% per-token discount. The $0.22 per task reconciles the two, and it is genuinely low; it is just not the 70% the release implies, because the release compares this model to the last MiniMax model.
What would change the verdict
Weights are the test. Community discussion points at October 23 for the download, while MiniMax says only later, so treat the date as expectation. When the files land the questions get concrete: what quantization survives a 1M context, what a four-bit build does at 214 GB of weights, whether the throughput advantage holds on hardware MiniMax does not own. Until then the fair summary is a cheap, capable, verbose model that is fast on someone else's servers and impossible to inspect.
