Introducing GPT-6 Sol and Luna

After launching GPT-6 Astra earlier this month as its top model, OpenAI is shipping GPT-6 Sol and GPT-6 Luna: cheaper, faster siblings trained with similar methods, aimed at everyday professional work, coding, and high-volume tasks.

API prices drop 50% versus GPT-5.6 promotional pricing — Sol to $2 / $10 per million input/output tokens, Luna to $0.10 / $0.50 — while OpenAI claims big gains on factuality and agent benchmarks.

We’re reading Sol and Luna as the distribution layer for Astra’s generation. Frontier models only change the world if mid-tier SKUs carry the same habits at prices teams can afford.

OpenAI says Sol takes difficult work with more room to iterate; Luna targets high-volume clerical goals — summaries, extraction, quick answers. On AutomationBench, Sol at xhigh effort beats Claude Opus 5 at max effort at about 9% of Opus’s cost per task; Luna gains 5.4 points over its predecessor at 58% lower cost. On Agents’ Last Exam, Sol at max hits 56.4%, above Opus 5’s best in that eval at 60% lower cost. Internal factuality evals — de-identified chats where users flagged prior mistakes — show Sol making about half as many mistakes as its predecessor, approaching Astra-level reliability cheaper. Coding and computer-use suites (FrontierCode, DeepSWE, OSWorld) are framed the same way: near-peer quality at a fraction of rival cost.

That’s useful AI if the numbers hold outside OpenAI’s charts. Lower cost plus better caching (90% off cached input reads, effort changes that don’t break cache) is how agents become default infrastructure instead of a CFO exception. Alignment evals also claim fewer misleading claims about coding work than GPT-5.6 peers.

The caveats are printed in the footnotes and we won’t ignore them: competitor scores from public reports, research/API evals that can differ from ChatGPT production, factuality sets biased toward previously flagged errors, AutomationBench notes on Fable fallback costs. Vendor leaderboards are advocacy. Still, a clean half-price cut is a market fact, not a vibes claim.

Availability: ChatGPT Work and Codex for paid tiers today; Luna for Free/Go on desktop; API as gpt-6-sol and gpt-6-luna; gradual ChatGPT rollout; not yet in Chat as of the post.

I’m watching independent evals and whether Sol actually steals default share from Opus/Fable on real workflows.

OpenAI announcement introducing GPT-6 Sol and Luna, expanding the GPT-6 family after Astra, with pricing, benchmark tables, caching notes, and rollout details.