MiniMax unveiled M3.1 Flash on October 19, 2026, claiming it would slash agentic task expenses by 70% and accelerate online inference to nearly three times M3's speed. The company priced Flash at $0.15 per million input tokens and $0.60 per million output tokens—half of M3's discounted rates of $0.30 and $1.20 for requests under 512K input tokens. However, hands-on testing against M3 showed the new model delivered on cost but fell short on performance metrics.

Artificial Analysis independently evaluated M3.1 Flash at 72.5% on AutomationBench-AA, a benchmark measuring multi-step task completion across simulated business applications. This metric reflects the proportion of task objectives accomplished without violating safety guardrails. The score differs from the full-task completion percentages tracked on Zapier's AutomationBench leaderboard.

To assess real-world performance, three separate agent tasks were run through both models via OpenRouter, routed to MiniMax's infrastructure at default settings. The testing methodology involved single runs per task through OpenRouter's free Flash endpoint, with cost estimates based on MiniMax's published pricing rather than actual charges. Direct MiniMax API testing was avoided due to prohibitively high credit minimums for a limited evaluation.

Test scenarios

  • Agentic bug fix: A Python job-queue library containing three interacting bugs and one incorrect test, with models reading, editing, and running tests via tools under a 25-turn limit.
  • Incident investigation: Logs, metrics, and deployment history from 20 simulated services during an outage, with the root cause obscured by two decoy deploys and a slow server clock.
  • Long context: A fabricated engineering wiki of approximately 500,000 tokens with three questions about service interdependencies.

Performance breakdown

Bug fix task

Both models achieved identical 5 of 6 scores but through different approaches. Flash resolved all three bugs without introducing errors but exhausted its turn allocation before documenting findings. M3 identified the wrong test but missed a case in the second bug.

Tool logs revealed contrasting strategies. Flash examined 37 files across its first 14 turns before making any edits, completing all bug fixes by turn 16 and spending remaining turns validating documentation. M3 executed tests immediately and applied its first fix on turn 4.

Flash's approach got all three bugs right, but it burned through its turns reading and checking and never got to write up what it found.

Flash consumed more than double the input tokens but achieved an estimated 23% cost reduction through cached input discounts. M3 processed faster at 118 tokens per second versus Flash's 55 tokens per second.

  • Flash: 25 turns, 538,718 input tokens, 7,056 output tokens, $0.017 estimated, 129.0 seconds, 55 tokens/s
  • M3: 17 turns, 203,639 input tokens, 4,327 output tokens, $0.022, 36.8 seconds, 118 tokens/s

Incident investigation task

Both models scored perfectly at 10 of 10, identifying the config change, eliminating both decoy deploys, detecting the slow clock, and correctly listing all nine affected services in sequence.

Despite making identical tool calls, Flash consolidated more calls per turn, finishing in 17 turns against M3's 24. Flash consumed fewer input tokens and achieved an estimated 68% cost reduction—closely matching MiniMax's stated 70% savings for agentic work. Turn count alone does not account for this difference; cache utilization and token pricing structures both contributed.

  • Flash: 17 turns, 367,401 input tokens, 13,175 output tokens, $0.020 estimated, 146.2 seconds, 90 tokens/s
  • M3: 24 turns, 592,377 input tokens, 12,061 output tokens, $0.062, 72.2 seconds, 167 tokens/s

Long context task

Neither model produced usable answers. Flash exhausted the 64,000-token output limit without generating a response. M3 similarly stopped after approximately 29,000 output tokens without completing the task. Despite MiniMax's specification of roughly 1-million-token context capacity for M3, this capability did not translate into task completion during testing.

MiniMax lists a roughly 1-million-token context window for M3, but that capacity did not translate into a completed answer…

A second attempt with reasoning set to low and a 28,000-token output ceiling yielded partial results. Flash produced an answer identifying only 3 of 33 affected services. M3 again ran out of tokens before responding. Neither model demonstrated reliable performance on large-document queries under these parameters.

Overall assessment

Both models matched scores on the two graded tasks, with neither completing the long-context question in initial runs. M3 demonstrated superior speed throughout testing. Flash's lower estimated total cost persisted despite higher input and output token consumption, a result influenced by cache mechanisms and differential pricing structures. Token counts alone cannot determine which model would prove more economical at identical rates.

Beyond the cost advantage, Flash did not demonstrate improvements in these test runs. Slower overall performance and identical accuracy scores on graded tasks suggest the cost reduction comes primarily from pricing rather than efficiency gains.

…apart from the lower estimated cost, I wouldn't call them improvements in these runs.

Three individual tests cannot validate whether MiniMax's benchmark claims generalize across broader workloads. Evaluating additional proprietary tasks before selecting either model would be prudent, with attention to MiniMax's forthcoming releases.