OpenAI ships GPT-6 Sol and Luna, cutting API prices in half
Sol nears Claude on coding at ~10% of the cost; Luna runs at $0.10/M in. Prompt-cache reads are now 90% off. Plus Qwen's 7B image model goes open.

Copy markdown
Sol and Luna, priced to replace your defaults
GPT-6 Sol drops to $2/$10 per million tokens (from $4/$20) and Luna to $0.10/$0.50 (from $0.20/$1.20) — permanent, not promotional pricing. Both slot in below the GPT-6 Astra flagship as cheaper everyday-work tiers.
Sol matches Claude on coding for a fraction of the spend
On DeepSWE v1.1, Sol at max effort scores 68.8% — within 1.1 points of Claude Fable 5's best — and on AutomationBench it beats Claude Opus 5's max-effort result at roughly 9% of Opus 5's cost per task.
Luna is the new floor for high-volume jobs
At $0.10 in / $0.50 out, Luna is aimed at summarization, extraction, and quick answers, yet still scores 66.6% on DeepSWE at about 93% less per task than Opus 5. Free and Go users get it in the desktop app.
Prompt caching just got much cheaper
Cached input-token reads now carry a 90% discount, and OpenAI says it cut the share of prompt tokens needing fresh processing by more than 50% across billions of requests — direct savings for repeated-context agent loops.
Live today in the API, Codex, and ChatGPT Work
Both models are callable now as gpt-6-sol and gpt-6-luna, rolling out in Codex and ChatGPT Work for Plus, Pro, Business, Enterprise, and Edu. Swap them into agent stacks and CI today.
Elsewhere: Qwen-Image-2.1 goes open with native transparency
Alibaba's 7B image model now outputs RGBA transparent-background images directly and edits from up to 10 reference images, with day-one ComfyUI, Diffusers, and vLLM support — but its research-only license blocks commercial use without a separate deal.