Short answer
MiniMax reports 56.22% on SWE-Pro, 55.6% on VIBE-Pro, and 57.0% on Terminal Bench 2 for M2.7. These are useful signals for software-engineering and agent-workflow evaluations, but they are vendor-reported: the captured materials do not provide full benchmark methods, dates, or independent replication. Treat the scores as reasons to investigate, not as proof M2.7 will win on your workload. (minimax.io)
For a new production API integration, however, M2.7 is no longer a practical shortlist candidate on MiniMax’s platform. MiniMax says new users lost access on October 19, 2026; M Plan and Token Plan requests using M2.7 are routed to M3.1 Flash; and existing pay-as-you-go users can use M2.7 only until its scheduled retirement on November 18, 2026. Existing users should plan migration rather than start a new dependency. (platform.minimax.io)
MiniMax M2.7 benchmark scorecard
The scores below are MiniMax-reported results, not independently confirmed rankings. Keep each score attached to its own benchmark: the tests use different tasks and metrics, so the numbers are not directly comparable to one another.
| Benchmark | MiniMax-reported result | What the source says | Important limitation |
|---|---|---|---|
| SWE-Pro | 56.22% | Software engineering; MiniMax says it nearly matches Opus’s best level on its product page. | No Opus score or matched evaluation conditions are supplied there. |
| VIBE-Pro | 55.6% | End-to-end project delivery. | No complete protocol or comparison results are supplied. |
| Terminal Bench 2 | 57.0% | Understanding complex engineering systems. | No detailed test setup is supplied. |
| GDPval-AA | ELO 1495 | Professional-work evaluation. | MiniMax’s product and GitHub pages use different wording for the “highest” comparison group (“open-source” and “open-weight”); avoid repeating that superlative as settled fact. |
| MLE Bench Lite | 66.6% medal rate across 22 ML competitions | Reported on MiniMax’s GitHub model page. | Vendor-reported; the page does not give a full protocol or independent replication. |
| SWE Multilingual; Multi SWE Bench | 76.5; 52.7 | Reported on MiniMax’s GitHub model page. | The captured text does not specify units for these two figures; do not add a percent sign. |
| NL2Repo | 39.8% | Reported on MiniMax’s GitHub model page. | Vendor-reported; detailed conditions are not supplied in the captured page. |
| Toolathon | 46.3% accuracy | Reported on MiniMax’s GitHub model page. | Vendor-reported; the page does not provide a detailed protocol. |
| MM Claw | 97% skill compliance across 40+ complex skills; 62.7% end-to-end | Reported on MiniMax’s GitHub model page. | The GitHub page’s 40+ skills description is not necessarily the same evaluation as the product page’s 40 cases described as over 2,000 tokens. |
The product page also reports 97% skill adherence across 40 complex-skill cases described as “>2000 Token.” MiniMax’s GitHub page separately reports 97% compliance across 40+ complex skills on MM Claw. The available descriptions do not establish that these are the same test, so keep them distinct. (minimax.io)
MiniMax also describes M2.7 as suited to software engineering, complex agent tasks, and office productivity, including multi-turn editing in Word, Excel, and PowerPoint. Those are vendor capability claims; they do not guarantee results in a particular application or production workflow. (minimax.io)
What the benchmarks mean for model selection
If you were evaluating M2.7 before its access change, the score categories would make it most relevant to software-engineering, project-delivery, ML, office-productivity, and tool-using agent workloads. They would not answer whether it is best for your codebase, prompts, tool harness, quality bar, or latency target.
A third-party review describes a Kilo Code comparison across three TypeScript codebases and reports both models finding planted bugs and security vulnerabilities, while noting differences in test coverage and code organization. That is the reviewer’s account, not a fully reproducible comparison in the material available here; it should not be treated as a standardized ranking. (platform.minimax.io)
A sound evaluation, whether you are assessing an existing M2.7 deployment or its successor, uses tasks from your own workload. Compare:
- Task success: Did the model complete the task and satisfy acceptance criteria?
- Output quality: Did code, edits, or agent actions pass your review and tests?
- Latency: Measure end-to-end response time under your actual prompt lengths and tool setup.
- Total cost: Include input, output, and any cache-billed tokens—not just the headline input rate.
- Operational fit: Confirm the model ID, route, response behavior, and migration requirements for the billing channel you use.
For example, take a set of representative coding or office tasks, run them through the model you currently use and the candidate replacement under the same instructions and tools, then score task completion and review effort. Record latency and actual token usage for each run. This is a proposed evaluation method, not a reported test result.
Production API: M2.7 access, migration, and price
M2.7’s API lifecycle matters more than its benchmark score now
MiniMax’s model-list notice says:
- October 19, 2026: M2.7 and M2.5 became unavailable to new users.
- M Plan and Token Plan: Requests using those model names are routed to M3.1 Flash starting October 19; the old names remain for backward compatibility.
- Pay-as-you-go API: Existing users may continue using M2.7 until November 18, 2026; after retirement, requests to the model will no longer work.
The same MiniMax documentation lists M2.7 in its retired-model section while its notice gives existing pay-as-you-go users the November 18 cutoff. For practical planning, follow the channel-specific notice: new users should not plan on M2.7 access, plan users should expect routing to M3.1 Flash, and existing pay-as-you-go users should migrate before November 18. (platform.minimax.io)
MiniMax’s M2.7 product page contains a minimal Python example using https://api.minimax.io/v1/text/chatcompletion_v2, model MiniMax-M2.7, and bearer-token authentication. That example documents the request shown on the page; it does not override the current access and retirement notice or establish current availability for a new integration. (minimax.io)
Current documented API alternative and rates
MiniMax’s current API overview lists MiniMax-M3.1-Flash and MiniMax-M3 as supported language API models. Its pricing page lists M3.1 Flash at $0.15 per million input tokens, $0.60 per million output tokens, and $0.015 per million prompt-cache-read tokens. Those are documented rates for M3.1 Flash—not confirmed M2.7 rates. The pricing page identifies token count as the billing item. (platform.minimax.io)
| Model/channel | What to know before choosing |
|---|---|
| M2.7, new API user | MiniMax says unavailable to new users from October 19, 2026. |
| M2.7, existing pay-as-you-go API user | Scheduled to stop working after November 18, 2026; plan a migration. |
| M2.7 name under M Plan or Token Plan | MiniMax says requests are routed to M3.1 Flash from October 19, 2026. The M Plan subscription key is separate from a pay-as-you-go API key. |
| M3.1 Flash, API | Current supported model; documented rates are $0.15/M input, $0.60/M output, and $0.015/M cache-read tokens. |
The API overview explicitly distinguishes pay-as-you-go API keys from M Plan subscription keys. Do not assume a subscription, an old model ID, or a price reported in a third-party review means M2.7 remains available at that rate through every access channel. (platform.minimax.io)
Bottom line
As a benchmark subject, M2.7 has notable vendor-reported results across coding, project delivery, tool use, and productivity. As a new production API choice on November 10, 2026, it is already closed to new users and near its scheduled pay-as-you-go retirement. Existing users should use the remaining period to run a representative comparison and migrate; new integrations should evaluate currently supported models, including M3.1 Flash, rather than rely on the M2.7 benchmark table alone. (platform.minimax.io)
Sources
MiniMax-M2.7 GitHub model page