OpenAI on September 22, 2026 broadened its GPT-6 family with two new models, GPT-6 Sol and GPT-6 Luna, priced at roughly half the tokens-per-dollar cost of last year's GPT-5.6 tier. Sol is aimed at complex professional work such as coding, agentic tool use and computer control; Luna targets high-volume clerical workflows.
Pricing And Reach
Sol drops to $2 per million input tokens and $10 per million output tokens (from $4 and $20). Luna is priced at $0.10 and $0.50 per million tokens. Both models ship inside ChatGPT Work and Codex for Plus, Pro, Business, Enterprise and Edu subscribers, and Luna is exposed to Free desktop users.
Head-To-Head Claims Aim Squarely At Anthropic
OpenAI says Sol matches Claude Fable 5 on the DeepSWE v1.1 coding DeepSWE test (68.8% vs 69.9%) at roughly 80% lower cost per task, matches Claude Opus 5 on OSWorld 2.0 computer-use scores at ~80% lower cost, and outperforms Opus 5 on OpenAI's own AutomationBench eval at just 9% of its cost per task. The company also claims Sol makes roughly half as many mistakes as its predecessor on internal factuality evaluations built from user-flagged errors.
Cheaper Compute, Same Frontier?
The launch turns up the pressure on rival frontier labs that just cut their own prices. Anthropic's Claude Opus 5.5 shipped the same week at 20% lower prices and 30% faster output, while OpenAI itself pushed GPT-6 Astra For Law earlier this month. Better prompt caching for GPT-6 also debuted alongside Sol and Luna.
Reporting based on OpenAI's launch materials and coverage from TechCrunch.