GPT-6 Sol and Luna halve API prices, rival Fable and Opus 5
OpenAI passes caching savings to the API: Sol matches Claude's coder at ~80% less per task, Luna nears Opus 5 for 10¢/M, as two Prime tiers hit OpenRouter.
Copy markdown
Sol: Fable-class coding at a fifth of the cost
GPT-6 Sol drops to $2/$10 per million tokens, down 50% from GPT-5.6 Sol. It hits 68.8% on DeepSWE v1.1 — within about a point of Claude Fable 5.1 — and 60.5% on OSWorld 2.0, edging Opus 5, at roughly 80% lower cost per task. Call it as `gpt-6-sol`.
Luna: near-Opus-5 quality for 10¢ per million
GPT-6 Luna is now $0.10/$0.50 per million tokens, down from $0.20/$1.20. It scores 66.6% on DeepSWE — comparable to Claude Opus 5 — and gains 5.4 points on AutomationBench over its predecessor at 58% lower cost per task. API id `gpt-6-luna`; Luna is also free to Free and Go users in the desktop app.
90% off cached input, live in Codex today
Both models ship with up to a 90% discount on cached input — often the bigger win for agents that reread the same context every turn. They're available now in ChatGPT Work and Codex for Plus, Pro, Business, Enterprise, and Edu users; OpenAI says improved caching and inference let it pass the savings straight through.
Two frontier Prime tiers land on OpenRouter
Also live this week: Qwen3.8 Max Prime (a 2.4T-parameter MoE with a 1M-token context) and Zhipu's open-weight GLM 5.3 Prime (753B MoE, 1M context, free to self-host) both appeared on OpenRouter — cheaper long-context options if you'd rather not lock into a single lab.