OpenAI expanded the GPT-6 family on September 22 with two new models, Sol and Luna, and the pricing is the headline: $2 per million input tokens for Sol and just $0.10 for Luna, with output at $10 and $0.50 respectively. Luna undercuts GPT-5.6’s promotional rates by half. Both sit below GPT-6 Astra, the flagship released earlier this month, and both inherit its training advances in coding, factuality, computer use, and alignment.
The price cut will dominate the coverage. Having spent the morning inside the rollout and the launch discussions, I’d argue the caching overhaul and a strange availability split deserve just as much attention.
Luna’s pricing changes what agents cost
Start with the numbers. Luna runs $0.10 per million input tokens and $0.50 per million output, which makes high-volume work almost embarrassingly cheap compared to a year ago. Sol, the stronger sibling, costs $2 in and $10 out. Both models handle up to 1.05M tokens of context, and both went live on Amazon Bedrock the same day, with full details in OpenAI’s announcement.
Sol’s internal evals show roughly half as many factual mistakes as GPT-5.6 Sol, alongside solid gains on coding benchmarks like DeepSWE and FrontierCode and on computer use tasks in OSWorld. Alignment testing also shows fewer misleading claims, which matters if you’re wiring these models into anything autonomous.
The quieter upgrade is prompt caching, and it’s the one I’d watch. Cached reads now cost 90% less, and adjusting reasoning effort or toggling tools mid-session no longer invalidates your cached context. That used to wipe the cache entirely. GitHub’s Copilot team is already seeing more than 50% less fresh token processing. For long-running agents, that saving compounds harder than the sticker price cut.
Plus subscribers still can’t use them in Chat
Now the friction. Sol and Luna run inside ChatGPT Work and Codex for Plus, Pro, Business, Enterprise, and Edu accounts. Free and Go users get Luna only, and only in the desktop app. Standard ChatGPT Chat gets neither. I watched the launch threads fill up within hours with paying subscribers asking why the new models aren’t in the interface they use every day.
Enterprise admins also have to enable the models manually, and OpenAI is phasing access in gradually for stability, with updates posted to OpenAI on X. If Sol hasn’t appeared in your workspace yet, that’s why. One genuinely useful quiet addition is that banked usage has been reset for Plus, Pro, and Business accounts, returning unused allowance from earlier cycles.
There’s a subtler change too. Both models inherited Astra’s communication style, which means shorter answers with less filler and jargon. In coding sessions it reads as a real improvement rather than a gimmick.
The Opus 5.5 collision
Anthropic released Claude Opus 5.5 minutes before OpenAI’s announcement. Same morning, deliberately or not. Early comparisons suggest Opus 5.5 still edges some raw benchmark scores, but OpenAI’s models win clearly on cost per completed task, especially agentic workloads where Sol’s caching keeps long sessions cheap. Benchmark graphs go stale in weeks. Pricing structures don’t.
This matters beyond developer circles. Cheap, always-on models are exactly what odds monitoring needs across sports betting markets.
The benchmark leaderboards matter less every quarter. What decides this race now is dollars per finished task, and OpenAI just moved the line. Watch two things: whether Sol and Luna ever reach standard Chat, and how long Anthropic waits before matching the cut. My bet is the price war turns ugly before the year is out.