In filings connected with xAI’s antitrust suit, OpenAI has said its Apple ChatGPT integration is underperforming and non-exclusive. That helps its defence. It also shows why GPT-6 Astra, priced at several times its predecessor, has to earn its money in enterprise workflows.
The most revealing thing OpenAI has said about distribution this month came in a court filing, not a product launch.
In a filing in xAI’s antitrust lawsuit, reported on September 23, OpenAI said that by January 2025 the Apple ChatGPT integration was “off to a slow start”, and later described it as “persistently underperforming”. It said it had cut its forecast for weekly active users gained through the integration, and it stressed, as Apple has, that the arrangement is non-exclusive.
Read on its own, that is an unusually candid admission about a flagship partnership. Read alongside the lawsuit, it is also a legal argument.
Why OpenAI wants the Apple deal to look small
xAI and X Corp filed the case in federal court in Fort Worth, Texas in August 2025, alleging that Apple and OpenAI colluded to lock up the smartphone AI-assistant market, with Apple favouring ChatGPT in iOS and its App Store at the expense of rivals such as Grok. The plaintiffs have since dropped their claims against Apple, but the case against OpenAI continues. It still depends on the partnership being large enough, and exclusive enough, to shut competitors out.
OpenAI’s best defence is to show that it is neither. An integration that underperformed internal forecasts and that Apple can extend to other model providers is hard to describe as a market-foreclosing arrangement. Every word about a slow start and non-exclusivity makes xAI’s harm harder to prove.
The admission is still evidence, since a lowered forecast is a specific claim made under the court’s scrutiny, but it is also advocacy.
Apple’s own choices support OpenAI’s account. In January Apple announced a multi-year, non-exclusive deal to build its next foundation models on Google’s Gemini, running on Apple’s Private Cloud Compute, and the Gemini-based Siri shipped with iOS 27 on September 14. Reports put the price at about $1 billion a year. Apple made Google, not OpenAI, the engine of its default assistant. That is the strongest evidence that the ChatGPT integration is neither exclusive nor central, and the best explanation for why it underperformed: iPhone users reach an assistant through Siri, and Siri now runs on Gemini.
Both readings point to the same commercial conclusion. Putting ChatGPT within reach of hundreds of millions of iPhones did not produce usage on the scale OpenAI expected. A user still has to enable the feature and choose to invoke it, and most apparently do not do so often. Access through an operating system is reach. It is not demand.
Astra is priced for demand
That lesson matters for OpenAI’s newest model, because GPT-6 Astra depends on demand.
Astra launched on September 3 and is being opened to selected organisations first and then to paid ChatGPT, enterprise, API and cloud customers. The standard API price is $10 per million input tokens and $50 per million output tokens, with a discounted batch tier and a premium fast tier. When GPT-5 launched in August 2025, its API cost $1.25 per million input tokens and $10 per million output tokens. Astra costs eight times as much to read and five times as much to write.
At that price, Astra has to do work that is worth far more than the inference bill. Coding, security analysis, complex research and multi-step enterprise tasks can support it if the model saves expensive human time or completes work cheaper models cannot. Occasional questions from a phone cannot.
The early customers fit that pattern. Airbnb is among the first enterprise users, reaching Astra through OpenAI’s APIs and Amazon’s Bedrock alongside other frontier models. Perplexity has described using Astra in end-to-end systems. These deployments give OpenAI something the Apple integration has not: metered usage attached to a business workflow, with a bill every month.
Astra’s risk profile points the same way. OpenAI classifies it as its first model to reach a “Critical” capability level for cybersecurity under its preparedness framework, which brings extra isolation and monitoring requirements. The model is most useful commercially in organisations already prepared to wrap it in enterprise controls. That describes large companies and cloud platforms, not a phone’s default assistant.
Where the price will be tested
Airbnb’s route to Astra through Bedrock matters here. On a cloud marketplace, Astra sits next to Anthropic’s and Google’s top models, and an enterprise can switch models with a configuration change. OpenAI cannot rely on default placement there, any more than it could on the iPhone. Astra has to keep winning workloads on quality per dollar, against rivals that can cut prices, every month.
That is the real test of the price. A premium of five to eight times the previous flagship is sustainable only if Astra completes tasks that cheaper models fail, not if it just does the same tasks a little better. The best early evidence will be retention. Enterprises that keep Astra inside expensive workflows after their pilots end, instead of routing most traffic back to cheaper models, will show that another step up in capability still commands a price.
Our view
The forecast figures behind OpenAI’s admission have not been made public, and later filings before Judge Mark Pittman may add more. They are unlikely to change the conclusion. With Siri running on Gemini, OpenAI has no default position on the iPhone left to defend. The Apple partnership will stay a low-cost source of consumer awareness and a minor contributor to revenue. OpenAI’s return on Astra will come from a few hundred large enterprise and platform customers, not from a billion devices. In court and in its pricing, OpenAI has already acted on that view.