OpenRouter report: open-weight models hit 60% of usage, ~70% of weekly tokens are Chinese models
Tiancaixinxin · x · 2026-10-08
Takeaways from OpenRouter's usage report at SF Tech Week:
- Open-weight models are taking over: share of total usage grew from 25% in January to 60% in September, and nearly all of them are Chinese models, making up 70% of weekly tokens.
- Agentic traffic drives cache rates: agents have higher cache rates and newer models are more optimized (>90%); blended average cost per token fell from $0.77 to $0.14 this year.
- Agentic traffic exceeded human traffic on Feb 16 this year.
- Labs didn't go agentic simultaneously — e.g., a 6-month traffic lag between Claude Opus and DeepSeek V4.
Related event: OpenRouter Report: Open-Source Models Hit 60% Share, Chinese Models ~70%(5 posts)→
More from Infra
- Reliquary opens its post-training network to Bittensor subnet teams, starting with batch data generation — const_reborn · 2026-10-08
- Samsung Q3: watch whether sales clear ₩200T — it would mean memory pricing runs hotter than contracts show — tengyanAI · 2026-10-08
- Google's Project Suncatcher starts testing TPUs in orbit for scalable AI compute — ZoubinGhahrama1 · 2026-10-08
- Claude Haiku 5.5 hits Databricks Day 0: ~15% better than Haiku 4.5 at a fraction of the cost — matei_zaharia · 2026-10-08
- Cloudflare ships cf sql query: SQL over all Cloudflare analytics in the terminal — irvinebroque · 2026-10-08
- Using AI-assisted math to fit an 11-Mac-mini home cluster on the smallest table — BLUECOW009 · 2026-10-08