OpenAI GPT-6 Sol and Luna launched on September 22 as lower-cost additions to the company’s newest model family. OpenAI cut their API prices by 50% from the promotional rates charged for the corresponding GPT-5.6 models, giving developers new options below flagship GPT-6 Astra.

 

Sol is aimed at demanding professional, coding and computer-use workloads, while Luna targets applications where very low unit cost matters. Both use training methods related to Astra, but OpenAI continues to position Astra as its strongest model for projects that require the highest available capability.

 

The launch sets three clear price points and product roles:

  • GPT-6 Sol costs $2 per million input tokens and $10 per million output tokens.
  • GPT-6 Luna costs $0.10 per million input tokens and $0.50 per million output tokens.
  • GPT-6 Astra remains the premium option for OpenAI’s most demanding workloads.

 

OpenAI GPT-6 Sol and Luna Cut API Prices

OpenAI’s published comparison puts Sol’s input rate down from $4 to $2 per million tokens and its output rate down from $20 to $10. Luna’s input price falls from $0.20 to $0.10, while output falls from $1.20 to $0.50 per million tokens.

 

The company attributes the lower prices to improvements in caching and inference. That distinction matters because the reductions are not presented as a limited discount on identical models; they accompany a new generation that OpenAI says improves capability while lowering the cost of serving each request.

 

OpenAI also upgraded prompt caching for GPT-6. Cached input reads can receive a 90% discount, and developers can change reasoning effort or tool availability without invalidating the earlier cached context. Applications with long conversations or recurring instructions may therefore see savings beyond the headline token rates.

 

Sol Targets Professional and Coding Work

GPT-6 Sol occupies the middle of the family: cheaper than Astra, but designed for work that still requires sustained reasoning and reliable tool use. OpenAI highlights business workflows, software engineering, factuality and operating computers as areas where Sol advances over GPT-5.6 Sol.

 

On OpenAI’s internal factuality evaluation, Sol made about half as many mistakes as its predecessor. The company says the test used de-identified conversations in which users had flagged earlier errors, so the result measures a deliberately difficult set rather than the frequency of mistakes in ordinary use.

 

OpenAI reports that Sol scored 68.8% on DeepSWE 1.1 at maximum reasoning effort, 1.1 percentage points below the highest Claude Fable 5 result in its comparison. It said the Sol run cost about 80% less per task, although customers will need to test those economics with their own prompts, tools and output lengths.

 

Further Reading

 

Luna Pushes Low-Cost Automation

Luna is the volume-oriented model. At one-twentieth of Sol’s input price and one-twentieth of its output price, it is positioned for high-frequency tasks such as classification, routine tool calls, lightweight coding assistance and customer-facing automation where every request must fit a tight cost envelope.

 

The lower tier is not simply a stripped-down legacy model. OpenAI says Luna inherits advances in professional work, coding, factuality, computer use and alignment from the GPT-6 training program. At maximum reasoning effort, the company reports a 66.6% score on DeepSWE 1.1.

 

Availability extends beyond the API. Sol and Luna began rolling out in ChatGPT Work and Codex for Plus, Pro, Business, Enterprise and Edu users. Free and Go users can access Luna in the desktop app, while OpenAI said the models were not yet available in Chat at launch.

 

Lower Inference Costs Reshape Model Selection

The release sharpens an important split in the model market. Developers no longer compare only which provider has the strongest benchmark result; they increasingly choose a capability tier for each step of an agent, routing difficult decisions to a premium model and routine work to cheaper ones.

 

Sol and Luna give OpenAI a broader answer to that routing strategy inside one model family. A coding agent might use Luna for file triage, Sol for implementation and Astra for the rare task requiring the deepest reasoning, subject to testing and the quality controls of the application.

 

OpenAI’s own benchmark claims remain vendor-reported, and lower token prices do not guarantee lower total costs. Reasoning effort, retries, tool calls, context size and the amount of generated output can dominate a production bill. Reliability and latency also determine whether a cheaper model is genuinely economical.

 

The immediate competitive pressure is nevertheless clear. With Sol and Luna available as gpt-6-sol and gpt-6-luna in the API, developers can test the new tiers without waiting for a future preview. Their adoption will show whether price, rather than peak benchmark leadership, becomes the decisive advantage for everyday AI work.