MUMBAI: The AI price tag is getting a trim, with OpenAI putting efficiency at the centre of its latest GPT-6 expansion. The company has launched GPT-6 Sol and GPT-6 Luna, cutting their standard API prices by 50 per cent against the promotional pricing of their GPT-5.6 counterparts.
GPT-6 Sol is priced at $2 per million input tokens and $10 per million output tokens, compared with $4 and $20 respectively for GPT-5.6 Sol. GPT-6 Luna costs $0.10 per million input tokens and $0.50 per million output tokens, compared with $0.20 and $1.20 for GPT-5.6 Luna. OpenAI’s current pricing page confirms the new standard rates.
The two models are positioned between OpenAI’s flagship GPT-6 Astra and the company’s previous generation, with Sol aimed at complex coding and agentic workflows and Luna designed for focused, repeatable, high-volume tasks. OpenAI says both models build on advances introduced with Astra, while improvements in inference and caching are intended to reduce the cost of running longer AI workflows.
For developers, the distinction is less about simply choosing a newer model and more about matching intelligence to workload. Sol is built for tasks such as developing features, reviewing code, debugging and analysing data, while Luna targets use cases including document summarisation, information extraction and quick questions.
OpenAI’s task-level figures underline that pitch. On AutomationBench, GPT-6 Sol at xhigh effort scored 33.2 per cent at a reported cost of $0.27 per task. Claude Opus 5 at maximum effort scored 26.9 per cent, with OpenAI reporting its cost per task as 11.1 times higher than Sol. GPT-6 Astra at low effort scored 30.3 per cent but cost 3.9 times as much per task as Sol.
On the Agents’ Last Exam, which assesses complex professional workflows, Sol at maximum effort scored 56.4 per cent. OpenAI said that result came at a 60 per cent lower cost per task than Claude Opus 5’s highest score in the evaluation.
Computer-use workloads also feature in the comparison. On OSWorld 2.0 offline, Sol at xhigh effort recorded a 60.5 per cent score, compared with 60.3 per cent for Claude Opus 5 at medium effort, while OpenAI reported approximately 80 per cent lower cost per task for Sol. Luna at maximum effort also exceeded GPT-5.6 Sol at medium effort at one-tenth of its cost.
Caching is another part of the cost equation. OpenAI said GPT-6 models can offer discounts of up to 90 per cent on cached input-token reads, allowing applications to reuse previously processed context instead of repeatedly paying the full input cost.
The company also pointed to the economics of its own AI usage. OpenAI said its daily token consumption, valued at API prices, exceeds $600 for the median researcher and $7,000 for researchers at the 90th percentile.
Lower prices could also make high-volume consumer and developer applications easier to scale. OpenAI cited Replit, which said the price reductions helped it introduce Free Mode to millions of users. The company also said GPT-5.6 Luna usage grew more than 10 times after an 80 per cent price reduction in July.
Coding performance has received another lift with Sol. OpenAI said the model substantially improves on GPT-5.6 Sol on FrontierCode, an evaluation focused on whether coding agents can produce changes ready to be merged into real-world codebases. The company also reported that Sol can match Claude Fable 5.1 at xhigh effort at a lower cost.
Factuality has also improved, according to OpenAI. On an internal evaluation using de-identified real-world conversations where users had flagged model errors, the company said GPT-6 Sol made about half as many mistakes as its predecessor. At higher effort levels, GPT-6 Luna can match GPT-5.6 Sol at around one-hundredth of its cost.
Both models also carry forward alignment improvements from GPT-6 Astra, with OpenAI reporting better results than their GPT-5.6 counterparts on alignment evaluations, including lower rates of misleading claims about coding work. The company cautioned that these tests are designed around challenging situations and should not be treated as typical-use failure rates.
GPT-6 Astra remains OpenAI’s highest-capability model for demanding workloads, while Sol and Luna are being positioned as lower-cost alternatives within the GPT-6 family. Both are available through the API as gpt-6-sol and gpt-6-luna.
The models have also rolled out in ChatGPT Work and Codex for Plus, Pro, Business, Enterprise and Edu users, while Free and Go users can access GPT-6 Luna through the desktop app.
The broader message from OpenAI is clear: as AI agents take on longer and more demanding jobs, the cost of every task matters almost as much as the capability of the model doing it.