OpenAI’s newest models arrived on Tuesday less than three weeks after the company’s flagship, and they made a simpler promise: nearly the same brain, half the price.
The company released GPT-6 Sol and GPT-6 Luna, two models that package the capabilities of its GPT-6 Astra system for different kinds of work. Sol is aimed at coding and other complex tasks; Luna is tuned for document summarization, information extraction and the high-frequency queries that dominate everyday use.
The economics are the headline. OpenAI said the API price for both models is 50 percent below the promotional pricing of GPT-5.6, a cut the company attributed to improvements in caching and inference efficiency. The discounts push the cost of running a near-frontier model down far enough to make it the default for a wide class of work that used to be priced as premium.
GPT-6 Astra, which OpenAI called the most capable model it has built, arrived less than three weeks ago and reset the top of the market’s expectations. The speed with which its abilities have now been pushed into cheaper models is itself a statement about how compressed the release cycle has become.
The models are available to Plus, Pro, Business, Enterprise and Edu subscribers through ChatGPT Work and Codex, and through the API. Free-tier and Go users can try Luna in the desktop app. Neither model has yet reached the standard ChatGPT chat interface, a sign that OpenAI is sequencing the rollout rather than putting its newest models in front of everyone at once.
Internally, OpenAI said Sol’s error rate on factual evaluations is roughly half that of the previous generation. The company has been pushing the idea that raw capability matters less than reliability, and the claim about error rates is aimed at the enterprise buyers who measure models by how often they are wrong, not how clever they are when right.
The timing was no accident. Anthropic released Claude Opus 5.5 about ninety minutes before OpenAI’s announcement, and the two companies spent the day trading price cuts and performance claims. The near-simultaneous releases have become a feature of the industry, as each lab tries to blunt the other’s news cycle.
The rivalry sits awkwardly alongside the joint posture the two companies struck earlier this month, when they each urged a slower approach to the frontier. The call was short on specifics and light on obligations, urging caution on the most advanced systems while saying little about the mid-tier models that generate most of the industry’s revenue. Tuesday’s releases, both below the frontier line, fit comfortably inside that gap.
OpenAI is framing Sol and Luna as a way to bring Astra’s advances down the cost curve, a pattern the company has followed since it began tiering its models. The flagship stays at the top, priced for the hardest problems, while the mid-tier models carry the volume. What is new is how quickly the flagship’s capabilities have flowed into the cheaper tiers, and how steep the discount has become.
OpenAI has run this play before. Its GPT-5 line was split into a family of models at different price points, and the 5.6 promotional pricing was an effort to hold off cheaper rivals. The new 50 percent cut moves the company from defending a price to setting a floor that others must now match.
The strategy is also a response to pressure from below. Anthropic, Google and a crowd of open-weight models have been cutting prices aggressively, and OpenAI’s promotional pricing on GPT-5.6 was itself an attempt to hold share. Cutting another 50 percent from that level signals that OpenAI would rather own the volume at thin margins than defend a higher price.
For developers, the practical question is which model to build on. Sol’s emphasis on coding positions it against tools aimed at software engineers; Luna’s focus on cheap, fast extraction positions it for the agent pipelines that read documents and fill forms. OpenAI is betting that splitting the models this way will let each win a distinct slice of spending.
The rollout leaves the company’s frontier story intact. Astra remains the system OpenAI calls its most capable, and the newer models do not displace it so much as extend it. The company’s bet, as with its rivals, is that the largest market is not the hardest problem but the millions of ordinary ones that run constantly and need to cost almost nothing.
Both models arrived, in other words, as products of the same arithmetic now governing the whole industry: capability is falling in price faster than anyone predicted, and the companies that survive will be the ones that can sell intelligence by the billions of calls rather than by the breakthrough.
