MiniMax released M2.5 on Friday, a Mixture-of-Experts model that activates just 10 billion of its 230 billion total parameters per inference pass. The company trained it with reinforcement learning across more than 200,000 real-world environments, and the model weights are available under a modified MIT license that requires commercial users to display the MiniMax M2.5 branding.
The pricing is the real story. M2.5-Lightning runs at 100 tokens per second and costs $0.30 per million input tokens, $2.40 per million output. That works out to roughly $1 for an hour of continuous operation. MiniMax claims this puts the model at one-tenth to one-twentieth the cost of Claude Opus 4.6, Gemini 3 Pro, and GPT-5, per their blog post. Four instances running nonstop for a year: $10,000.
On company-reported benchmarks, M2.5 scores 80.2% on SWE-Bench Verified and 51.3% on Multi-SWE-Bench. Independent testing from OpenHands ranks it fourth overall, behind Claude Opus models and GPT-5.2 Codex, but calls it the first open-weight model to exceed Claude Sonnet. "Intelligence too cheap to meter," per MiniMax's own framing, though the benchmarks they highlight are largely self-reported or run on their own infrastructure.
MiniMax says 30% of internal tasks at the company are now handled by M2.5, with 80% of newly committed code generated by the model. Hong Kong-listed shares jumped 15.7% on the news. The API and a standard 50 TPS variant (at half the Lightning price) are live on minimax.io.
The Bottom Line: M2.5 is the strongest open-weight coding model yet by independent measures, and its per-token pricing could make always-on AI agents economically viable for the first time.
QUICK FACTS
- Architecture: MoE, 230B total parameters, 10B active
- SWE-Bench Verified: 80.2% (company-reported)
- M2.5-Lightning: 100 TPS, $0.30/1M input, $2.40/1M output
- Standard M2.5: 50 TPS, $0.15/1M input, $1.20/1M output
- License: Modified MIT (requires branding for commercial use)
- MiniMax HK shares: +15.7% to HK$680 on release day
