MiniMax M2.5 — Near-Frontier Agents at a Fraction of Opus Economics
MiniMax released M2.5 and M2.5 Lightning (Feb 12): vendor SWE-Bench Verified 80.2%; Lightning ~100 tok/s at $0.30/$2.40 per MTok. Open variants on Hugging Face (modified MIT); positioned for continuous agents ~$1/hr in company estimates. Cost-per-task is the trade—not another Elo screenshot.
When a model can sustain high-quality agentic work for roughly $1 per hour, new application classes become practical—persistent agents, always-on monitoring, mass parallel runs. MiniMax’s proprietary frame: sell cost-per-task, not only bench proximity to Opus.
MiniMax launched M2.5 and M2.5 Lightning on February 12. Vendor scores: 80.2% on SWE-Bench Verified; leading or near-leading on Multi-SWE-Bench and BrowseComp (company post). Focus: coding, agentic tool use, search, office productivity; trained with RL across hundreds of thousands of complex environments (company). Lightning: 100 tokens/second (2× typical frontier throughput cited), $0.30/M input and $2.40/M output; standard variant even lower priced; both support caching. Company/coverage: ~1/10–1/20 the cost of Claude Opus 4.6 / peer frontier output pricing. Open variants on Hugging Face under modified MIT; full API also offered. Lightning targets speed-critical deployments with identical capability claim vs standard.
Claims vs checks
Pricing, speed, and license claims are MiniMax primary. SWE-Bench / BrowseComp figures are vendor—cross-check independent leaderboards. VentureBeat and others relayed cost comparisons; treat multiples as approximate, not audited TCO. “$1/hour” continuous-agent estimate is company-framed arithmetic.
Limits
- Cheap tokens ≠ cheap end-to-end agents once tools, retries, and human review enter.
- Modified MIT: read redistribution clauses.
- Throughput claims depend on serving stack and batch size.
Sources
- MiniMax: “MiniMax M2.5: Built for Real-World Productivity” (February 12, 2026). Primary.
- VentureBeat and independent coverage on cost comparisons and benchmark proximity (Feb 12, 2026).
- Hugging Face model cards / community throughput reports.