OpenAI has launched GPT-6 Sol and GPT-6 Luna, two models aimed at bringing parts of the GPT-6 capability family to more workloads and budgets. OpenAI says they are available in ChatGPT Work and Codex, with API model IDs gpt-6-sol and gpt-6-luna.
The practical difference is model choice: Sol targets more demanding professional and coding tasks, while Luna is the lower-priced option for routine work and higher-volume applications. OpenAI says GPT-6 Astra remains its strongest model overall, so Sol and Luna are alternatives for balancing capability, speed, and cost rather than replacements for every task.
Last verified: September 22, 2026, using OpenAI’s launch announcement and developer documentation. Product access is rolling out, so availability can vary by account and API project.
GPT-6 Sol and Luna at a glance
| Model | API input price per 1M tokens | API output price per 1M tokens | OpenAI’s positioning |
|---|---|---|---|
| GPT-6 Sol | $2.00 | $10.00 | More demanding professional and coding work |
| GPT-6 Luna | $0.10 | $0.50 | Lower-cost everyday tasks and applications |
OpenAI compares these prices with the promotional pricing of GPT-5.6 Sol and Luna and says the new models reduce API prices by about 50%. The listed Luna output rate falls from $1.20 to $0.50 per million tokens, which is a larger reduction. Prices are per million tokens and do not include the full operating cost of an application, such as tools, retries, or human review.
What OpenAI announced
OpenAI describes Sol and Luna as more affordable members of its GPT-6 model family, bringing improvements it attributes to Astra across professional work, factuality, coding, computer use, and alignment. The company also points to changes in inference and prompt caching as part of the efficiency work behind the launch.
For developers, the announcement adds two API choices: gpt-6-sol and gpt-6-luna. OpenAI says prompt caching has higher default cache-hit rates for GPT-6, offers a 90% discount on cached input-token reads, and now allows reasoning effort and tool availability to change without invalidating earlier cached context. Explicit cache breakpoints give developers additional control over which prompt prefix is reused.
These are OpenAI’s product and performance statements. The company’s benchmark results are not a substitute for testing a model against your own data, tools, and success criteria.
Where Sol and Luna may fit
GPT-6 Sol for complex work
Sol is the option to evaluate when a task needs longer reasoning, coding, multi-step tool use, or professional workflow completion. OpenAI reports results on its selected benchmarks, including coding and business workflow tests. Those results depend on benchmark design, model settings, and task mix; they do not predict every team’s production results.
GPT-6 Luna for cost-sensitive volume
Luna’s listed token rates make it a candidate for applications where the previous model’s price limited experimentation or volume. Lower token prices can make it easier to test more model-assisted steps, but they do not make a workflow automatically inexpensive. Measure total cost per accepted result, including retries, tool calls, latency, and correction time.
Keep Astra for tasks that need its higher capability ceiling
OpenAI says GPT-6 Astra remains its best model across the board. A useful architecture can route difficult or high-impact requests to Astra while using Sol or Luna for tasks that meet a defined quality threshold at lower cost. The route should be based on evaluation data and include a safe fallback for low-confidence results.
Availability and API model names
OpenAI says GPT-6 Sol and Luna are available in ChatGPT Work and Codex for Plus, Pro, Business, Enterprise, and Edu users. Free and Go users can access GPT-6 Luna in the desktop app. OpenAI says the models are not yet available in Chat, and that a gradual rollout may mean they do not appear immediately in every eligible account.
For API use, the published model IDs are gpt-6-sol and gpt-6-luna. Confirm that the model appears in your project and review the official model documentation before updating an application. Availability and supported features can change during rollout.
What the pricing means for application teams
OpenAI’s listed rates are a meaningful change for teams that run many model calls or reuse long prompt prefixes. Caching can reduce the cost of repeated context, while lower input and output rates can make it practical to evaluate a wider set of tasks.
For a fair comparison, track:
- cost per task that meets your acceptance criteria;
- quality and factual error rate on representative examples;
- latency, including tool execution and retries;
- cache hit rate and cached versus uncached input;
- the amount of human correction required;
- safe fallback behavior when the model is uncertain or a tool fails.
Benchmark comparisons in a launch announcement should be read as results under the stated test conditions. OpenAI itself describes limitations in some of its evaluations, and comparisons across providers may use different effort settings, fallback behavior, and cost accounting. Run a small controlled evaluation before changing production routing.
Limitations and responsible rollout
Lower prices do not remove the need for quality and safety controls. Sol and Luna can still make mistakes, misunderstand instructions, or use tools in ways that require oversight. For applications that can affect customer data, finances, access, or production systems, use scoped permissions, logs, human approval for consequential actions, and a rollback path.
Do not select a model based only on a benchmark headline. A smaller model that needs repeated retries or extensive corrections may cost more in practice than a stronger model that completes the task reliably. Compare the models on the actual workflow and keep the test set representative as the application changes.
How I can help you evaluate GPT-6
I help teams plan and build AI features through AI consulting and custom AI development. That can include model selection, API integration, evaluation plans, caching, cost controls, and review steps for tool-using applications.
For a complete product, I can also support SaaS product engineering, workflow automation, and AI features for website development or mobile app development. The right starting point is a bounded test with a clear quality target and a cost per completed task.
Official sources
Frequently asked questions
What are GPT-6 Sol and GPT-6 Luna?
GPT-6 Sol and GPT-6 Luna are OpenAI models launched on September 22, 2026. OpenAI positions Sol for more demanding professional and coding work, while Luna offers a lower-cost option for everyday tasks and applications.
What are the GPT-6 Sol and Luna API model IDs?
OpenAI lists the API model IDs as gpt-6-sol and gpt-6-luna. Check the official model documentation and your API project for current availability and supported features before changing production traffic.
How much do GPT-6 Sol and GPT-6 Luna cost through the API?
OpenAI lists GPT-6 Sol at $2 per million input tokens and $10 per million output tokens. GPT-6 Luna is listed at $0.10 per million input tokens and $0.50 per million output tokens. These are token prices; tool use, retries, and application overhead can add to total workflow cost.
Who can use GPT-6 Sol and Luna?
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. The models are not yet available in Chat, and API availability may depend on rollout and project access.
Are GPT-6 Sol and Luna always better than GPT-6 Astra?
No. OpenAI says GPT-6 Astra remains its strongest model overall. Sol and Luna are positioned as more cost-efficient choices; teams should compare quality, latency, and total cost on their own tasks before selecting a model.