Arena.ai says that MiniMax-M3.1-Flash entered the open-model rankings at No. 12 and the overall rankings at No. 32 in Code Arena: WebDev with a score of 1588. That is 106 points higher than MiniMax-M3’s 1482. This is a snapshot of human preference rankings for WebDev tasks and should not be treated as a measure of general model capability.
Tasks it can help assess: Relative preference signals for web development, frontend generation, and WebDev Agent workflows.
Tasks it should not be extrapolated to: General chat, knowledge Q&A, non-WebDev software engineering, production stability, end-to-end success rates, or real-world performance in other clients.
Applicable model versions: MiniMax-M3.1-Flash, with MiniMax-M3 as the comparison. The post does not provide a more specific model snapshot or reasoning level.
Applicable client, Agent, or API: Arena’s Code Arena: WebDev evaluation environment. The result cannot be directly restated as a score for the MiniMax API, MiniMax Code, or another Agent harness.
Reasoning level and parameters: Not specified.
The original post only says that the result comes from Arena’s Code Arena: WebDev and gives the model rankings, scores, previous-generation comparison, and category changes. It does not disclose the task set, questions or prompts, sample size, number of votes, score aggregation formula, model snapshot, harness, tool configuration, reasoning level, margin of error, or confidence interval. The 1588 score therefore cannot be reproduced, nor can the statistical stability of adjacent rankings be assessed, from the original post alone.
The post also reports changes in the overall ranking and in the rankings of seven WebDev categories. The category score differences are comparisons stated directly in the Arena post; they cannot be used to infer the number of tasks, success rate, or independent score for each category.
| Metric | MiniMax M3.1 Flash | Comparison in the post |
|---|---|---|
| Code Arena: WebDev | 1588 points; No. 12 among open models; No. 32 overall | — |
| MiniMax lab ranking | No. 12 among all labs | — |
| Previous-generation model | — | MiniMax-M3: 1482 points; No. 19 among open models; No. 66 overall |
| Change in score relative to previous generation | +106 points | M3.1 Flash relative to MiniMax-M3 |
| Price | $0.15 input / $0.60 output per 1M tokens | The post says M3.1 Flash costs 50% less |
Category ranking and score changes:
Reference-Based Design: No. 20 → No. 7 (+155 points).
Content Creation Tools: No. 20 → No. 7 (+113 points).
Data & Analytics: No. 18 → No. 8 (+68 points).
Gaming: No. 21 → No. 9 (+179 points).
Simulations: No. 22 → No. 13 (+122 points).
Brand & Marketing: No. 18 → No. 14 (+78 points).
Consumer Product: No. 18 → No. 14 (+64 points).
The original post explicitly gives 1588 pts, #12 among open models, and #32 overall.
The original post lists MiniMax-M3 as 1482 pts, No. 19 among open models, and No. 66 overall, and says that M3.1 Flash is +106 pt higher.
The original post gives $0.15 for input and $0.60 for output, both per 1M tokens, and says that the cost is 50% lower; it does not provide the cost calculation sample, token composition, or previous-generation price baseline.
The original post gives the before-and-after rankings and score differences for seven categories; see “Key results.”
The original post was published on 2026-10-19; the figures in this article are a snapshot of the post collected on 2026-10-21.
The supported conclusion from this source is that, in the relevant release snapshot of Arena Code Arena: WebDev, MiniMax-M3.1-Flash scored 1588, ranking No. 12 among open models and No. 32 overall, while its publicly reported comparison score was 106 points higher than MiniMax-M3’s. Arena also reports ranking and score improvements for it in all seven WebDev categories.
These results can only serve as human preference signals for a specific set of WebDev tasks and voting window. A WebDev human preference score is not equivalent to general capability, and it cannot support claims that the model is more reliable on other tasks, toolchains, or production projects. The original post omits the benchmark method and sample size, so the statistical significance of the differences and the complete reproducibility of the leaderboard cannot be assessed.
“50% lower cost” is a claim made in the Arena post. Because it does not disclose the previous-generation price baseline, request composition, token ratio, or cost sample, it cannot be converted into actual savings for an arbitrary business.
Open https://x.com/arena/status/2112156275282968930, enter the post details, and click “Show original” to confirm the body and the 2026-10-19 timestamp.
Fix the WebDev evaluation snapshot from the post, recording the 1588 score, the two overall leaderboard rankings, MiniMax-M3’s 1482 score and rankings, and the changes in the seven categories. Do not mix later dynamic leaderboard figures into this record.
For an independent rerun, separately obtain Arena’s task set, voting sample, scoring rules, model snapshot, harness, tool permissions, and reasoning configuration, and label the new result as an independent rerun.
MiniMax M3.1 Flash
