MiniMax M3.1 Flash Released: Agent Task Scores Reported at Top; API Cost and Time-of-Day Billing Are Key Developer Variables
According to a Xinzhiyuan report dated October 19, 2026, MiniMax released M3.1 Flash, which is live on MiniMax API, M Plan and MiniMax Code, with the official CLI's default text model switched. The report says it ranks at or near the top on Automation-Bench and LOCA-Bench 256K, among others, and comes with 1M-token context and five effort levels.
What changed
According to a Xinzhiyuan report dated October 19, 2026, MiniMax officially launched M3.1 Flash two days after the release of DeepSeek V4.1 Flash-1017, and the report says it is already live on MiniMax's official API, M Plan and MiniMax Code, with the default text model of the official CLI repository switched to MiniMax-M3.1-Flash. The report says the model natively supports 1M-token context and can directly read text, images and video; thinking mode is on by default, and an effort parameter can adjust across five levels—low, medium, high, xhigh and max—defaulting to max when unspecified. Benchmark data relayed by the report: DeepSWE v1.1 scores 73.7 at the max level, Automation-Bench 62.8 for first place, LOCA-Bench 256K 60.3, BankerToolBench 76.2 and CyberGym 82.2; on the computer-use side, OSWorld 2.0 scores 61.7 and Agents' Last Exam has a 29.6% pass rate, with the report estimating a full ALE run costs about 739 yuan. The report also says everyday output speed reaches 200+ TPS with a median stable above 170+ TPS, and that its predecessor was the anonymous model 'Space Bunny'. Note: all the above are official or third-party benchmark data relayed by the report, not independently verified here, and some comparison model names and versions are beyond what can be verified.
What it means for developers
The impact on developers centers on three things: billing structure, default parameters and context strategy. The report says M3.1 Flash costs 1 yuan per million input tokens, 4 yuan for output and 0.1 yuan for cache reads, with no time-of-day pricing; it also contrasts this with DeepSeek V4.1 Flash-1017, which the report says prices peak and off-peak in two tiers—2 yuan input and 8 yuan output during weekday daytime peak hours. If that description holds, enterprise Agent workloads that mainly run during the day would have the cost gap between the two amplified by time-of-day differences, while overnight batch workloads would see a much smaller gap. The second point is the effort parameter's default: the report says it defaults to max when unspecified, meaning integrations that don't pass it explicitly consume thinking tokens and latency at the heaviest level, raising both bills and response times. Third, 1M-token context and direct multimedia reading expand the input a single request can carry, turning 'fill in the whole thing or retrieve and trim' into a trade-off that needs real-world cache-hit testing; the report's 1-billion-input-token estimate reaches different conclusions at 90% versus 95% cache-hit rates. Fourth, the report says the official API documentation lists M3.1 Flash first among models, and that a Pro version aiming for the intelligence ceiling comes later, with no timeline or availability commitment yet.
Integration and evaluation
The following are suggestions and conditional analysis based on the report, not verified facts. Decision one: whether to make M3.1 Flash the default model for Agent workflows. The basis is the report's 62.8 on Automation-Bench for first place and roughly 0.25 to 0.3 yuan per task. The actual decision point is the effort level—not passing it explicitly means running at max throughout. Applicable conditions: batch processes spanning multiple SaaS systems where overall success rate matters more than per-step latency can be swapped in for trials first; single-step, low-risk, latency-sensitive calls should be explicitly lowered to low or medium. Evaluation standard: build a fixed set of 50 to 100 tasks from real business processes, run each once at medium and once at max, and record end-to-end success rate, total tokens, actual per-task cost and P95 latency, with the pass condition being success rate no lower than the current model and per-task cost lower than the current setup. Decision two: whether long context and multimedia input replace the existing retrieve-and-trim pipeline. The basis is the reported 1M-token context and direct text, image and video reading. The decision point is filling in everything versus keeping trimming, and the condition is the actual cache-hit rate—the report shows different cost conclusions at 90% and 95% hits. Evaluation standard: measure both pipelines on the same evaluation set for answer consistency, first-token latency and P95 latency, and cost per thousand calls converted at the real hit rate, then take the one with acceptable quality difference and lower cost. Decision three: whether to mix routing among Chinese Flash models by task type. The report says DeepSeek is stronger on DeepSWE software-engineering tasks while MiniMax stands out on enterprise workflows and financial tool calling, which is the report's summary rather than a settled conclusion. The premise is that the two APIs' error semantics, timeouts and retries can be wrapped uniformly. Evaluation standard: total success rate and total cost of the same mixed task batch under a single model versus the routing scheme; if the added complexity of routing doesn't buy a measurable improvement in success rate or cost, fall back to a single model.
Limits and open questions
To be clear: as of this catalog check, MiniMax M3.1 Flash does not exist in KimiSeek's current model list, nor does any MiniMax-family entry, so the reported API availability refers to MiniMax's official channels; this article does not state that the model is available on KimiSeek. All benchmark scores, speeds, costs and price descriptions come from the Xinzhiyuan report of October 19, 2026 and are relayed official or third-party evaluations, not independently reproduced here; some comparison model names and versions are beyond what can be verified and should not be used to infer an actual capability ranking. The reported figures such as 1 yuan per million input tokens and 4 yuan for output reflect the reporting at the time of release and do not constitute current pricing; before integrating, refer to MiniMax's official pricing page and documentation, and this article does not restate or lock in any current price. The report says a Pro version comes later, which is forward-looking information with no release date or availability commitment; the quality and cost differences across effort levels are also only described at the report level and need your own testing. Real throughput under different regions, concurrency and rate limits may differ notably from the reported TPS figures.