MiniMax M3.1 Flash vs. M3: Lower Costs, Mixed Results in Agent Tests

Oct 26, 2026 379 views
Abstract 3D render of coin-like discs between flowing rainbow chrome waves, evoking AI token pricing

MiniMax M3.1 Flash’s clearest advantage in my small comparison with M3 was its estimated price. Across three one-off agent tasks, Flash cost less, but it didn’t improve the scores on the two tasks I could grade, and it took longer overall. For teams weighing model costs against task completion time, that’s a useful distinction: a lower per-token price doesn’t automatically mean a cheaper or faster workflow.

MiniMax released M3.1 Flash on October 19, 2026. The company says the model reduces the overall cost of agentic tasks by 70% and makes online inference nearly three times faster than M3. Its listed prices are $0.15 per million input tokens and $0.60 per million output tokens. M3’s discounted rates for requests up to 512K input tokens are $0.30 and $1.20 respectively; cached input and longer requests have separate pricing.

Those figures describe token prices and inference speed, not necessarily the time an agent takes to finish a job. A tool-using model may read files, make calls, check its work, and produce a final response. Each step affects the total time and token bill.

What the benchmark can—and can’t—tell us

Artificial Analysis gave M3.1 Flash a score of 72.5% on AutomationBench-AA, which tests multi-step work in simulated business applications. The score represents the proportion of task objectives completed without guardrail violations. It isn’t the same measure as the share of whole tasks completed on Zapier’s AutomationBench leaderboard.

I didn’t run that benchmark. Instead, I compared Flash and M3 on three smaller jobs that resemble practical engineering work: fixing a buggy Python project, investigating a simulated outage, and answering questions from a large technical wiki. Each model got one run per task, with a cap on the number of turns. That sample can show how these runs differed; it can’t establish a general advantage in accuracy, speed, or cost.

Test setup

Both models ran through OpenRouter, routed to MiniMax’s servers, with default settings. Flash was available through OpenRouter’s free endpoint during the tests, so its dollar figures below are estimates using MiniMax’s listed prices, not charges I paid. I used OpenRouter to run both models in one place rather than MiniMax’s API, where the credit minimum was too high for this small comparison. Results could differ through MiniMax’s API, even when requests reach the same servers.

  • Agentic bug fix: A Python job-queue library contained three interacting bugs and one incorrect test. Each model could inspect and edit files, run tests, and use other tools, with a 25-turn limit.
  • Incident investigation: Logs, metrics, and deployment history covered 20 fictional services during an outage. The underlying cause was a configuration change obscured by two decoy deployments and a server clock that ran slow.
  • Long context: A fictional engineering wiki of about 500,000 tokens supplied the source material for three questions about dependencies among its services.

Bug fixing: careful inspection, different omissions

Both models finished with five of six points, but for different reasons. Flash corrected all three bugs without causing regressions. It ran out of turns before writing a final response, however, so it didn’t report the incorrect test. M3 identified that test but missed one case in the second bug.

The traces show why their results diverged. Flash spent its first 14 turns reading 37 files before editing. By turn 16, it had fixed the bugs, then used its final four turns to write a scratch test and check the documentation. M3 ran the tests and made its first fix on turn four. Flash’s thoroughness helped its code changes, but it left no room to explain the test problem.

Flash used more than twice as many input tokens as M3 in this task, yet its estimated cost was about 23% lower, helped by discounted cached input. It also took much longer to run. Here’s the task-level accounting:

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

Incident response: Flash used fewer turns and tokens

The outage investigation was the strongest result for both models. Each scored 10 out of 10, found the configuration change, dismissed both decoy deployments, caught the slow server clock, and named all nine affected services in the right order.

They made the same 63 tool calls, but Flash grouped more calls into each turn and completed the task in 17 turns. M3 needed 24. Flash also used fewer input tokens, though it produced slightly more output. Its estimated cost came to about 68% less than M3’s for this run, close to MiniMax’s stated 70% reduction for agentic tasks. That match is suggestive, not confirmation: this was one task, and token prices and cache use shaped the bill alongside the turn count.

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

Long context: neither model completed the task

The large wiki exposed a limitation that a context-window figure alone can’t resolve. Neither model answered the three questions in the first run. Flash consumed the full 64,000 output tokens allowed in the test without producing an answer. M3 stopped after roughly 29,000 output tokens and also failed to answer. MiniMax lists an approximately one-million-token context window for M3, but fitting a large document into context didn’t lead to a useful result here.

I made one further attempt with reasoning set to low and an identical 28,000-token output limit for both models. Flash returned an answer, but identified only three of the 33 affected services. M3 again used its output allowance without answering. Under these settings, neither model handled the dependency question reliably; Flash’s shorter answer missed most of the requested services.

Results across the three tasks

TestMiniMax M3.1 FlashMiniMax M3
Agentic bug fix5 of 65 of 6
Incident investigation10 of 1010 of 10
Long contextNo answerNo answer
Total tokens in / out1,401,606 / 84,2311,291,477 / 45,214
Total cost (Flash estimated)$0.15$0.27
Total model time661.6 s465.0 s

The aggregate figures capture the trade-off. Flash’s estimated cost was $0.15, compared with $0.27 for M3, while its total model time was 661.6 seconds against M3’s 465.0. Flash processed more tokens in total, but token counts alone don’t explain the price difference because the models have different rates and cache use can affect input costs.

What these runs suggest

My read is that M3.1 Flash looks promising when the goal is to reduce spend on work that benefits from batching tool calls, as the incident investigation did. The bug-fix run shows the other side: Flash spent more time inspecting and checking, then ran out of turns before it could deliver a complete report. On the large-document task, neither model produced a dependable answer.

These three runs don’t show whether the benchmark result generalizes, or whether Flash will be cheaper across a team’s own workloads. Both models earned the same scores on the two graded tasks, and M3 finished the combined runs sooner. Teams considering the price reduction should compare the models on representative jobs, including the final written output and end-to-end completion time—not just inference speed or token rates. The next useful evidence will come from repeated tests across varied tasks and settings.

The post MiniMax M3.1 Flash vs. MiniMax M3: One is cheaper, but not faster or better appeared first on The New Stack.

Source: Jessica Wachtel · thenewstack.io

Comments

Sign in to comment.
No comments yet. Be the first to comment.

Related Articles

MiniMax M3.1 Flash vs. MiniMax M3: One is cheaper, but no...