OpenAI's GPT-5.6 tiers go global as Chinese open-source models capture 46% of enterprise traffic
GPT-5.6 Sol, Terra, Luna hit global public access while Chinese models undercut US pricing by 60-90%.
OpenAI opened GPT-5.6 Sol, Terra, and Luna to global users after a two-week preview. Sol scores 88.8% on Terminal-Bench 2.1 (91.9% in Ultra mode with orchestrated subagents) and targets biology, chemistry, and cybersecurity. Pricing per million tokens ranges from $5/$30 for Sol down to $1/$6 for Luna. The tier structure allows enterprises to route tasks: Terra approximates GPT-5.5 quality at half the price, while Luna suits high-volume classification. GPT-5.6 also introduces explicit prompt cache breakpoints and a 30-minute minimum cache life.
Meanwhile, CNBC reported Chinese open-source models now account for 30-46% of US enterprise API token traffic, sustained above 30% since February 8. GLM 5.2 from Z.ai saw 80x customer growth and 27x daily token volume in its first week on Vercel. These models run 60-90% cheaper than US frontier models and land within a point of Opus 4.8 on agentic benchmarks. Anthropic moved Claude Cowork to cloud with web and mobile clients, enabling persistent sessions across devices even when offline. The company also published J-Space research, a Jacobian-based interpretability tool that reveals Claude's silent internal reasoning, with practical applications for auditing and safety review.
- OpenAI's GPT-5.6 tiered pricing ranges from $1/$6 per M tokens (Luna) to $5/$30 (Sol), with Sol achieving 88.8-91.9% on Terminal-Bench 2.1.
- Chinese open-source models now handle 30-46% of US enterprise API traffic; GLM 5.2 had 80x customer growth and costs 60-90% less than US models.
- Anthropic expanded Claude Cowork to cloud/web/mobile with cross-device persistence and released J-Space, an interpretability tool to inspect silent reasoning.
Why It Matters
Model routing is now a competitive advantage—overspending on frontier models for routine tasks can be avoided with cheaper alternatives.