OpenAI’s newest flagship model does more than its predecessor. It also costs more. The company said it is pricing GPT-6 Astra’s application-programming interface at $10 per million input tokens and $50 per million output tokens, with cache reads and writes billed separately, a step up from the $5 and $30 rates that GPT-5.6 Sol commanded when it launched.
The increase is a direct statement about how OpenAI values its most advanced work. Frontier models have historically followed two pricing trajectories: falling as chips get cheaper and systems get more efficient, or rising when a model is judged to be in a class of its own. With Astra, OpenAI is betting on the second trajectory, setting the price high enough that the gap between it and the previous generation reads as a feature rather than an accident.
Developers who build on OpenAI’s platform will feel the change where they live: in the cost per request. An application that summarizes documents or powers a chat interface pays on every call, and a doubling of input prices plus a 67 percent increase on outputs changes the unit economics of products that were priced against the old model. Startups running high-volume workloads will have to decide whether Astra’s gains justify the premium, or whether Sol remains good enough for the jobs it already does.
The price also signals how OpenAI thinks about its customer base. The flagship tier is not aimed at the developer who needs cheap completions. It is aimed at enterprises and builders who need the most capable reasoning available, the kind of customers who pay for quality because the cost of a wrong answer exceeds the cost of a more expensive right one. For that group, a difference of a few dollars per million tokens is a rounding error against the value of a model that can operate a computer or reason through a security problem.
Caching, the other revenue line, is where the real money sits for heavy users. Models that remember prior turns of a conversation avoid reprocessing the same text, and OpenAI has long charged separately for reading from and writing to that cache. The company did not publish the Astra cache rates alongside the headline numbers, but the decision to call them out separately suggests caching will be an increasingly important part of how customers manage cost on long-running agents.
The pricing arrives with the broader rollout of the model, which OpenAI released this week to its Daybreak cybersecurity customers before spreading it to paid tiers and the API over the following days. Astra is the model whose reasoning OpenAI has acknowledged is harder for outsiders to audit, and the company’s chief scientist has said its “opaque looped reasoning” makes its thinking process less visible to humans. Paying more for a model whose reasoning is less transparent is a trade some buyers will accept and others will question, and the API price now makes that trade explicit in dollar terms.
Industry analysts said the move reflects a market that has room for both strategies. Closed frontier labs still command premiums for their most advanced systems, while open-weight models have been undercutting them on price for routine work. OpenAI’s tiered pricing lets it occupy both ends: cheap enough on older models to compete with open alternatives, and expensive enough on the flagship to fund the compute bill that training the next generation requires.
There is a precedent for the strategy inside OpenAI’s own history. Each major model release has arrived with its own price ladder, and the company has generally reserved the top of the ladder for reasoning models that spend more compute per answer. Astra extends that logic: a model that can browse the web, operate software and plan over long horizons consumes far more resources per task than a model answering a single question, and its price partly reflects that physics.
The comparison that matters to developers is not Astra against Sol on paper, but Astra against whatever they were running before. A product that used to send every request to a $5 model and now sends it to a $10 one has just doubled its largest variable cost. OpenAI is betting that the quality gap justifies the switch, and the early enterprise reception, including Microsoft’s decision to make Astra available through Azure to its customers within days, suggests the company believes the demand is there.
Pricing is also a form of positioning against rivals. Anthropic, Google and xAI all sell frontier access, and each release cycle brings the same question: who sets the ceiling on what the best model costs? By pricing Astra above its predecessor, OpenAI is claiming that its latest model is worth more than the generation that preceded it, a claim that holds only if the model’s real-world results justify it.
For now, the market will answer in usage data. If developers route their workloads to Astra in volume despite the higher rates, the price will look like a correct read of the model’s value. If they stay on Sol, or drift to cheaper open alternatives, OpenAI will face the usual pressure to adjust. Either way, the $10 and $50 numbers settle one thing: the company believes its most capable model is also its most expensive to run, and it is asking customers to pay for what that costs.
