OpenAI halves GPT-6 API prices with Sol and Luna
OpenAI has split its GPT-6 family in two, launching Sol and Luna with the abilities of the Astra generation and API prices half those of the models they replace.
OpenAI expanded its GPT-6 line on Tuesday with two models built for everyday work rather than the top of the benchmark table. GPT-6 Sol and GPT-6 Luna were trained with the same methods as GPT-6 Astra, released earlier this month, and carry its advances in professional tasks, factual accuracy, coding, computer use and answer style at a fraction of the price.
The headline change is cost. OpenAI has halved API prices against the promotional pricing of the GPT-5.6 generation: Sol falls from $4 to $2 per million input tokens and from $20 to $10 per million output tokens, while Luna drops from $0.20 to $0.10 on input and from $1.20 to $0.50 on output. GPT-6 Astra remains the company’s most capable model, and OpenAI pitches the new pair as the cheap end of the same family rather than a replacement for it.
Where Sol and Luna land
Both models are available from today in ChatGPT Work and Codex for Plus, Pro, Business, Enterprise and Edu users. Free and Go users get Luna in the desktop app, the two models are not yet in Chat, and developers can call them as gpt-6-sol and gpt-6-luna. OpenAI says it will roll the pair out gradually through the day, so the models may take a few attempts to appear.
What the benchmark table claims
OpenAI’s own numbers put Sol close to the frontier on professional tasks and coding while undercutting its rivals on cost per task. On AutomationBench, which tests agents on end-to-end business tasks across 47 tools, Sol at extra-high effort scored 33.2% at $0.27 a task, ahead of Claude Opus 5 at max effort on 26.9% at more than eleven times that cost. On Agents’ Last Exam, which evaluates long-horizon professional work, Sol reached 56.4% at max effort, above Opus 5’s highest score at 60% lower cost per task. On DeepSWE v1.1, Sol scored 68.8% against Claude Fable 5’s 69.9% for roughly 80% less per task, with Luna on 66.6%. OpenAI also says Sol makes about half as many factual mistakes as its predecessor on its internal evaluation of flagged conversations.
Cheaper context for agents
Alongside the price cut, OpenAI has reworked prompt caching so agents reuse more context by default, with a 90% discount on cached input reads. Developers get a dashboard that shows how much input is being cached, a diagnostics tool for missed caching opportunities and explicit breakpoints that control where a cached prefix ends. Changing reasoning effort or tool availability no longer clears the cache. GitHub says the changes have cut the share of prompt tokens needing fresh processing by more than 50% across billions of requests to OpenAI models, which is one reason Copilot answers faster.
Our opinion
The frontier labs have quietly changed what they are competing on. OpenAI is not selling Sol and Luna as the smartest models available, and its own page concedes Astra remains the best model on the board; it is selling them as the cheapest models that are nearly as good, with a comparison table aimed squarely at Claude Opus 5 and Fable 5.1. For anyone running agents rather than chatting, that is the metric that matters, because cost per completed task decides how much work you can hand over. The caveat is that every figure is self-reported, and the two evaluations buyers care about most, long-horizon coding and computer use, are precisely where OpenAI still says it trails. Halving a price is only a saving if the workload does not simply grow to fill the budget.