OpenAI launches GPT-6 Sol and Luna, lowering inference and programming costs and upending AI agent economics

OpenAI has launched the GPT-6 Sol and GPT-6 Luna models, expanding the GPT-6 family with faster, lower-cost versions tailored for professional and programming tasks and for running independent agents. The two new models build on the technical advances introduced by GPT-6 Astra, but they bring those capabilities to a broader operational scope at lower prices and compute consumption, benefiting from direct improvements in inference efficiency and temporary memory management of input data.

The new pricing reduces the API cost by fifty percent compared with the promotional rates for the GPT-5.6 model.Token consumption cost has dropped to half across the APIs.Sol costs $2 per million input tokens and $10 per million output tokens, down from $4 and $20 respectively, while Luna’s cost fell to $0.10 for inputs and $0.50 for outputs per million tokens, down from $0.20 and $1.20. The company also offers a discount of up to ninety percent on temporary-storage reads, along with monitoring and diagnostic dashboards that let developers analyze data reuse and interpret retrieval failure cases.

In terms of performance on programming tasks and agent benchmarks, Sol achieved 33.2 percent on the Automation Bench at ultra-high effort with a cost of $0.27 per task, surpassing GPT-6 Astra at low effort and Cloud Opus 5. In the Agents Last Exam evaluation, Sol scored 56.4 percent at maximum effort, outdoing the best result of Opus 5 while reducing cost per task by sixty percent. On the Deep SW E 1.1 code-writing metric, Sol reached 68.8 percent, just 1.1 points shy of the top score recorded by Cloud Fibel 5, but at roughly eighty percent lower cost per task.

Internal measurements showed the synthetic-code deception rate falling to 1.3 percent for Sol and 2.8 percent for Luna, compared with about 10.4 percent for the previous GPT-5.6 Sol. Real-error occurrences for Sol also dropped to roughly half of the prior generation in an internal test that used conversations flagged for errors, noting that the test samples were not representative of typical usage and were not adjusted for answer length. By contrast, Luna posted an improvement of 5.4 percentage points over its predecessor under high-effort conditions, while reducing task-completion cost by fifty-eight percent. GitHub reported that caching improvements cut the tokens needing fresh processing by more than half across billions of programming requests.

The new models are available via the API under the identifiers gpt-6-sol and gpt-6-luna, and the company has begun offering them in the ChatGPT Work and CodeX environments for Plus, Pro, business, enterprise and education subscribers, with Luna free for desktop-app users and Sol provided to all Perplexity engine users, while both models remain outside the standard public chat application so far.

This steep drop in inference prices and the large temporary-memory discount has a direct impact on software development teams and companies in the Gulf, Egypt and the Levant.Changing the compute-cost equation opens the door to building AI agents at a truly commercial scale.Operating budgets had been a barrier to deploying complex, recurring processing tasks such as continuous database auditing or building programming assistants for banks and startups. An eight-percent reduction in cost for programming tasks gives local technical teams the ability to run automated testing systems and iterate code without facing escalating inference bills, and it prompts technology leaders in the region to reassess the architecture and costs of their applications, leveraging advanced software efficiency at prices that were previously limited to lightweight models.