OpenAI Halves API Prices With GPT-6 Sol and Luna as Anthropic Cuts Opus Running Costs 40%

OpenAI released GPT-6 Sol and GPT-6 Luna on September 22, nineteen days after GPT-6 Astra, and priced both below the models they replace. Luna costs $0.10 per million input tokens and $0.50 per million output tokens. Sol costs $2 and $10. Both prices are half the current promotional rates for the equivalent GPT-5.6 models. The batch application programming interface (API) cuts those already halved prices by another 50 percent, which puts Luna's batch output price at $0.25 per million tokens.

Hours earlier, Anthropic introduced Claude Opus 5.5. The company says it performs at the level of Claude Fable 5.1 on most work while costing about 40 percent less to run than Opus 5. The list price fell 20 percent, to $4 per million input tokens and $20 per million output tokens. Anthropic attributes the rest of the savings to two factors: cache reads, which account for most agentic and coding costs, dropped 60 percent, and the model uses fewer tokens to complete a task.

The benchmark results are mixed, and both companies' numbers come from their own testing. OpenAI reports that Sol scored 33.2 percent on AutomationBench at its highest reasoning setting, against 26.9 percent for Claude Opus 5, at roughly one-ninth the cost per task. On DeepSWE, a software engineering benchmark, Sol scored 68.8 percent and Luna 66.6 percent, below Claude Fable 5's 69.9 percent but at about 80 percent lower cost. Anthropic itself acknowledges that at this level of capability, benchmark differences are a weak guide to real differences in production use. A Russian-language technical channel summarizing the OpenAI launch described Sol and Luna as built on GPT-6 Astra's work but tuned for faster, cheaper operation at scale, a fair characterization of the release.

OpenAI paired the launch with a rebuilt prompt-caching system: a 90 percent discount on cached input tokens, cache breakpoints that let developers choose which prefixes persist, a diagnostics dashboard, a 30-minute reuse window, and the ability to change reasoning effort between responses without invalidating the cache. OpenAI says GitHub has cut the share of prompt tokens processed from scratch by more than half across billions of Copilot requests.