OpenAI is extending its GPT-6 generation with new Sol and Luna models, emphasizing lower prices and fewer mistakes rather than another top-tier capability claim. The launch follows GPT-6 Astra earlier this month and fills out the company’s model hierarchy for work that does not require its most powerful option.
Sol is positioned for complex tasks such as coding. Luna targets high-volume work with a clear objective, including document summaries, information extraction and quick answers. That separation gives customers a choice between a heavier problem-solving model and a faster model designed for routine throughput.
Half-Price Access Is the Main Commercial Shift
OpenAI says API access to the GPT-6 versions of Sol and Luna will cost half as much as the corresponding 5.6-series models. The company attributes the reduction to improvements in caching and inference, although TechCrunch does not provide the token prices or an independent cost analysis.
The pricing change matters because a model used repeatedly for clerical or coding tasks can accumulate costs through volume rather than a single expensive request. Luna’s focus on clearly specified, repeated work makes the efficiency claim particularly relevant to organizations processing many documents or questions.
OpenAI also says the models produce more factually accurate answers and make fewer coding errors. In an internal factuality evaluation based on de-identified conversations in which users had flagged mistakes, GPT-6 Sol reportedly made about half as many errors as its predecessor and reached Astra-level reliability at a lower cost.
That result is useful but bounded. It comes from OpenAI’s own evaluation and from a selected set of conversations associated with reported mistakes. The source does not provide a complete benchmark table, sample size or independent replication, so the “half as many” figure should not be generalized to every task.
Competitive Claims Arrive Minutes Apart
OpenAI repeatedly compares the new models with Anthropic’s Fable and Opus tiers, saying Sol and Luna perform substantially better on several tasks. TechCrunch notes that Anthropic released Opus 5.5 only 90 minutes before OpenAI’s announcement.
The close timing turns two separate releases into a visible contest over capability, serving cost and model positioning. Anthropic lowered the price of its highest Opus tier, while OpenAI is presenting smaller GPT-6 models as a route to Astra-like reliability and broader affordability.
Both sets of performance claims originate with the companies selling the models. The article does not supply a neutral comparison across identical workloads, and the vendors describe different roles for their tiers. A headline benchmark lead therefore does not establish which model will be cheaper or more dependable inside a specific coding or document workflow.
Distribution Reaches Work, Developers, and Consumers
The new Sol and Luna versions are available through ChatGPT Work and Codex for most paid accounts, as well as through the ChatGPT API. Luna is also coming to the desktop application and to Free and Go users. OpenAI said the wider ChatGPT rollout would proceed gradually during the day.
That distribution gives the launch several tests at once. Developers can measure price and behavior through the API, professional users can apply the models to coding and office tasks, and a broader consumer audience can encounter Luna through lower-cost access.
The release does not eliminate the need to match a model to a task. Sol’s complex-work positioning and Luna’s volume orientation imply different tolerances for latency, cost and error. OpenAI’s internal results suggest progress, but organizations will still need their own evaluations for the documents, codebases and decisions that matter to them.
GPT-6 Astra established the high end of the generation. Sol and Luna may have the larger operational effect if their lower prices make repeated use practical. Whether they deliver that effect depends less on launch-day comparisons than on how reliably the efficiency and factuality claims survive real workloads.
