MiniMax shipped M3.1 Flash as the generally available version of its M3.1-Flash-Preview, aimed at agentic coding, computer use, finance tooling and security tasks. The newsroom held the story because the only verified source, the API overview, listed the model ID, a 1M-token context window and adjustable thinking depth. It did not support the claims circulating with the launch: the date, the price, the speed, free access and open weights.
This matters to builders because the price is the lowest MiniMax lists for an LLM. It also comes with a long context window, which makes it a plausible default for high-volume agent loops. But a cheap, fast model that confidently makes things up can cost more downstream than it saves. The useful question is which launch claims hold up against primary and independent sources.
What we checked
- Can a primary MiniMax announcement or model-specific pricing page verify the release date, price, speed, and access claims? partly confirmed. MiniMax's launch blog (dated 2026-10-19) and its API pricing page confirm the date and price: $0.15 input, $0.60 output and $0.015 cached input per 1M tokens. They also confirm access through the API, M Plan and MiniMax Code, plus the speed claims (200+ tok/s, P50 170+, 'nearly 3x faster'). Neither page mentions free access or gives a date for open weights. The blog only says weights 'will be released later'. [1][2][3]
- Does independent measurement support MiniMax's 200+ tokens/s speed claim? refuted. Artificial Analysis lists output speed at 104.7 tokens/s on its model page. That is about half the vendor figure and slightly below the median for its price tier. Different setups (the first-party endpoint versus a benchmark harness, and max reasoning effort) could explain part of the gap. [4][1]
- Are the weights openly available now? refuted. Artificial Analysis lists the model as proprietary with no public weights. MiniMax's blog says the weights will come 'later'. VentureBeat and The Decoder report a planned date of October 23, but I could not confirm that date on a MiniMax page I could open. [4][1][6][7]
What we found
- The release and price are confirmed by MiniMax itself. The launch blog is dated 2026-10-19 and gives flat per-1M-token rates of $0.15 input, $0.60 output and $0.015 cached input. The API pricing page lists the same rates for MiniMax-M3.1-Flash with no long-context tier. MiniMax-M3, after a 'permanent 50% off', is $0.30/$1.20 up to 512k input tokens and doubles above that, with a Priority tier at 1.5x. [1][2]
- MiniMax's speed claims (200+ tokens/s in everyday use, P50 170+ tokens/s, 'nearly three times faster' online inference) are the company's own. Artificial Analysis measured 104.7 tokens/s, slightly below the median for its price tier. The '70% lower usage cost' figure compares agentic-task totals with the previous generation. MiniMax's tools page says it is not a comparison of API unit prices. [1][3][4]
- Access is through the MiniMax API, the M Plan subscription and the MiniMax Code desktop app. MiniMax's pricing page lists no free allowance. OpenRouter hosts a 'MiniMax M3.1 Flash (free)' variant, but the page does not say how long it lasts or what the limits are, beyond rate-limit (429) handling. [1][2][5]
- MiniMax's own benchmark figures are BankerToolBench 76.2, DeepSWE v1.1 73.7%, OSWorld V2 61.7, Agents' Last Exam 29.6 and CyberGym 82%. The blog compares them only against unnamed 'top models'. Independently, Artificial Analysis scores the model 43 on its Intelligence Index (launch coverage said 42) and puts it #4 of 173 in its class. It costs about $0.22 per Index task, which VentureBeat reports is down from about $0.51 for M3. [1][4][6]
- Reliability is the main caveat. VentureBeat and The Decoder, citing Artificial Analysis, report a 93% hallucination rate on AA-Omniscience. The model attempts nearly every question and rarely says it doesn't know. It answered about 56% correctly but scored only 16 overall on that index. [6][7]
- The weights are not open. Artificial Analysis lists the model as proprietary. MiniMax's blog promises the weights 'later' with no date. A planned October 23 date appears in VentureBeat and The Decoder, but no MiniMax page I opened confirmed it. Search snippets mention a 428B-total, 23B-active MoE design, but neither the blog nor MiniMax's tools page states a parameter count. [4][1][6][3]
What it means
The launch is real, and you can trust the price: $0.15/$0.60 with cheap cache reads and no surcharge for long prompts makes MiniMax-M3.1-Flash a strong candidate for high-volume agent loops, coding sub-agents and long-context batch jobs. MiniMax's tools page says reasoning is always on, with five effort levels defaulting to the most expensive. Budget for those reasoning tokens. Artificial Analysis's roughly $0.22 per Index task was measured at max effort. For latency planning, assume something closer to the independently measured ~105 tokens/s than the vendor's 200+ until you've benchmarked your own traffic.
The hallucination profile is the deciding factor. Don't use it unchecked for fact retrieval, finance or other high-stakes autonomous actions. Pair it with retrieval grounding, tool-verified outputs, or a verifier or second model on critical steps, and test whether lower effort settings change how often it declines to answer. Don't plan self-hosting around October 23 until MiniMax actually publishes weights and a license. Treat OpenRouter's free endpoint as a short-lived trial, not a production tier.
Still open
- Will MiniMax publish the weights on October 23 as reported, and under what license?
- What are the actual architecture and parameter count (the 428B/23B-active figure is unverified in primary sources)?
- Why does MiniMax's 200+ tok/s claim differ from Artificial Analysis's 104.7 tok/s measurement (endpoint, effort level, load)?
- How long does OpenRouter's free variant last, and with what rate limits?