MiniMax M2.5 Open-Weight Model Hits 80.2% SWE-bench at 1/20th the Cost of Claude Opus 4.6

VentureBeat·high signal

MiniMax released M2.5 and M2.5 Lightning, open-weight models using a 230B/10B-active MoE architecture that scores 80.2% on SWE-bench Verified (vs. Opus 4.6's 80.8%) and 51.3% on Multi-SWE-bench (vs. Opus 4.6's 50.3%), effectively matching frontier coding performance. At $0.15 per SWE-bench task completion versus $3.00 for Opus 4.6, M2.5 delivers near-identical coding quality at 5% of the cost, with MiniMax reporting 80% of internal code submissions now generated by M2.5. The open-weight release makes this the most cost-efficient near-frontier coding model available and directly undercuts every proprietary API for agentic code generation workloads.