OpenAI has introduced GPT-6 Sol and GPT-6 Luna , two faster, less expensive members of the GPT-6 family. GPT-6 Astra remains the model for the hardest work. Sol and Luna are meant to bring more of that generation’s capability to tasks that run often enough for cost to matter.
That distinction is important for agents. A coding assistant may read thousands of lines, call tools repeatedly, test several approaches, and revisit the same context. A model’s price per token matters, but so do its success rate, the number of attempts it needs, and how much context the application can reuse.
The lineup and the prices
OpenAI lists the following API prices per one million tokens in its launch announcement:
| Model | Input | Output | Likely role |
|---|---|---|---|
| GPT-6 Sol | $2.00 | $10.00 | Difficult work at a lower cost |
| GPT-6 Luna | $0.10 | $0.50 | High-volume tasks and inexpensive agent steps |
Sol’s listed input and output prices are half the promotional prices for GPT-5.6 Sol. Luna’s input price is also halved; its listed output price falls from $1.20 to $0.50, a larger reduction. Those are token prices, not a promise that every workflow’s total bill will fall by the same percentage.
The practical question is which model finishes a task reliably at the lowest cost per completed task. Sol may be a sensible default for substantial coding or business workflows. Luna is worth testing for frequent, simpler steps and for tasks where its lower price creates room for more iterations. Astra remains the option when a failed attempt or a missed detail is especially costly.
Stronger agents at a lower cost
OpenAI reports improvements across professional workflows, coding, and computer use. A few results show the scale of its claim:
| Evaluation | GPT-6 result | Comparison reported by OpenAI |
|---|---|---|
| AutomationBench | Sol, xhigh: 33.2% at $0.27 per task | Claude Opus 5, max: 26.9% at about 11 times Sol’s cost |
| DeepSWE 1.1 | Sol, max: 68.8% | Claude Fable 5, xhigh: 69.9% at roughly five times Sol’s cost |
| DeepSWE 1.1 | Luna, max: 66.6% | Comparable to Opus 5 and Fable 5 at medium effort in this evaluation |
| OSWorld 2.0 offline | Sol, xhigh: 60.5% | Claude Opus 5, medium: 60.3% at roughly five times Sol’s cost |
These figures come from OpenAI’s launch post. Effort settings differ, and the results cover specific agent setups and benchmark versions. They are useful starting points for testing, not a universal ranking of models. In particular, a one-point gap on a coding benchmark tells a team less than whether a model can make a change that passes its own repository’s tests and review standards.
The Luna result is notable because it suggests a low-priced model can now handle some long-running engineering tasks that once required a much more expensive tier. Whether that holds in a particular codebase depends on the task mix, tools, context, and tolerance for rework.
Factuality and communication are part of the cost equation
OpenAI says Sol makes about half as many mistakes as GPT-5.6 Sol on its internal factuality evaluation. Luna improves too, and at higher effort reportedly reaches GPT-5.6 Sol’s reliability at a much lower cost. The evaluation uses conversations where users had already flagged a factual error, so its error rates should not be read as typical production failure rates.
The company also says Sol and Luna inherit Astra’s clearer collaboration style: shorter answers, less jargon, and more direct accounts of what the model did. That is more than a writing preference for an agent. A precise status update helps a reviewer see what changed, what was checked, and where the agent remains uncertain. Confident but vague reporting can hide expensive follow-up work.
OpenAI reports improvements in alignment evaluations as well, including fewer misleading claims about coding work. Those tests deliberately create difficult situations and do not measure how often such behavior occurs in ordinary use. Teams should still verify changes with tests, diffs, and clear permissions.
Caching makes long sessions cheaper
The other major change is prompt caching. OpenAI says GPT-6 now has higher cache hit rates by default and offers a 90% discount on cached input-token reads. Developers can change reasoning effort or tool availability without invalidating earlier cached context, and can use explicit breakpoints to choose where a reusable prompt prefix ends.
For a long coding session, that could make repeated repository instructions, prior messages, and tool context much cheaper to reuse. OpenAI says GitHub has seen the share of prompt tokens requiring fresh processing fall by more than 50% across billions of requests to OpenAI models. That is a report about GitHub’s workload, not a guaranteed saving for every application.
Caching changes how teams should measure model economics. Compare complete workflows with realistic context, tool calls, retries, and cache hit rates. A tiny isolated prompt will miss much of the benefit that matters in an actual agent session.
Where Sol and Luna are available
OpenAI says both models are available in ChatGPT Work and Codex for Plus, Pro, Business, Enterprise, and Edu users. Free and Go users can access Luna in the desktop app. At launch, the models are not yet available in Chat. OpenAI says rollout in ChatGPT Work and Codex is gradual.
Developers can use the API model IDs gpt-6-sol and gpt-6-luna.
The larger shift is that frontier capability is spreading across price points. Astra raises the ceiling; Sol and Luna make more capable agents economical for repeated work. The best choice will come from measuring completed tasks in your own environment, including failures, review time, and cache savings.
Top comments (0)