OpenAI Apple ChatGPT Deal Fell Short, Exposing the Limits of Built-In AI

OpenAI says the OpenAI Apple ChatGPT deal underperformed despite placing its chatbot inside one of the world’s most valuable consumer technology platforms.

A September 22 court filing says the integration started slowly after its December 2024 launch. OpenAI reduced its forecast for additional logged-in weekly active users by January 2025. The actual figures remain redacted.

That admission reverses the logic surrounding the partnership. Apple offered extraordinary distribution through Siri, Writing Tools, and other system features. Yet distribution did not automatically produce active ChatGPT users, subscriptions, or the expected brand lift.

The disclosure comes from OpenAI’s motion for summary judgment in an antitrust case brought by Elon Musk’s X and AI company, now named SpaceXAI. The plaintiffs argued that Apple’s arrangement unfairly favored ChatGPT and restricted Grok’s access to iPhone users.

OpenAI’s response is unusually blunt. Its defense says the partnership was nonexclusive, reached relatively few users, and failed to create the market power alleged by its rival.

That makes this more than a disappointing partnership update. It offers a rare look at how an AI company measured the value of operating-system distribution behind closed doors.

It also shows the conflict between reach and adoption. A service can appear inside millions of compatible devices without becoming part of users’ daily behavior.

What the OpenAI Apple ChatGPT Deal Actually Delivered

The partnership provided privileged placement, but users still had to make several deliberate choices before ChatGPT became useful inside Apple’s software.

Apple and OpenAI signed their service integration agreement on May 4, 2024. Apple publicly introduced ChatGPT support alongside Apple Intelligence on June 10.

The original pitch was broad. ChatGPT would answer requests that Siri could not handle, work with documents and images, and support Apple’s systemwide writing features.

Users would not need a ChatGPT account for basic access. Existing subscribers could connect their accounts to use additional features available through those subscriptions.

Apple also placed privacy controls around the handoff. Siri would ask permission before sending a request, document, or image to ChatGPT. Apple said users’ IP addresses would be obscured, while OpenAI would not store unauthenticated requests.

Those decisions made sense within Apple’s privacy model. They also introduced friction into a product whose value depended on repeated use.

The integration went live on December 11, 2024, according to OpenAI’s court filing. It was turned off by default.

Enabling it required a multistep opt-in process through Siri or the Settings application. Apple’s current setup instructions still direct users through several screens before they can activate the extension and connect an account.

Even after activation, ChatGPT did not replace Siri. Apple’s assistant decided when an outside model might help, or users explicitly requested ChatGPT.

That distinction matters. A native app invites users to open a recognizable destination with a clear purpose. An assistant integration asks users to understand a routing system that often remains invisible.

OpenAI monitored both direct and indirect benefits. Direct effects included usage and subscriptions attributed to Apple Intelligence users. Indirect effects included promotion, brand association, and activity through other channels.

Internally, OpenAI called the second category a “halo effect.” The company expected Apple’s endorsement to increase ChatGPT’s visibility and encourage activity beyond the integration itself.

By January 2025, that expectation was already weakening. The filing says the rollout appeared to be off to a slow start, prompting a lower forecast for incremental logged-in weekly active users.

The new forecast is redacted. So are several internal assessments that might explain the size of the disappointment.

However, the direction is clear. OpenAI revised its expectations only one month after launch. By summer 2025, the company says the integration was dramatically underperforming.

The problem persisted. The filing refers to the Apple integration as “persistently underperforming” during communications dated March 25, 2026.

This evidence does not show that nobody used the feature. It shows that usage fell below OpenAI’s internal expectations, including expectations revised shortly after launch.

Apple’s original Apple Intelligence announcement emphasized access without switching between tools. In practice, reduced app switching was not enough to establish a strong acquisition channel.

The result separates availability from adoption. ChatGPT was present inside Apple’s software, yet OpenAI says too few users converted that presence into measurable, recurring engagement.

Why ChatGPT Siri Adoption Started So Slowly

The central adoption problem was not access to ChatGPT. It was the distance between an available capability and a habit users understood.

Apple’s integration had several layers of friction. Users needed compatible hardware, supported software, an eligible language and region, and Apple Intelligence itself.

They then had to activate the ChatGPT extension. Connecting an OpenAI account required another decision, while some requests introduced confirmation before information left Apple’s environment.

Each step protected user control. Together, those steps reduced the chance that someone would discover ChatGPT accidentally and continue using it.

The experience also divided responsibility between Apple and OpenAI. Siri remained the visible interface, while ChatGPT operated as an outside service invoked for selected tasks.

That structure weakened OpenAI’s direct relationship with the user. Someone receiving an answer through Siri might not view the interaction as a reason to open ChatGPT later.

Logged-in weekly active users were especially important because OpenAI could identify their continued activity. Anonymous requests routed through Apple offered less direct evidence of customer acquisition.

The partnership therefore had two possible jobs. It could make Siri more capable, or it could bring users into OpenAI’s product ecosystem.

Those jobs were related but not identical. A successful Siri response might satisfy an iPhone owner without producing an OpenAI login, subscription, or repeat visit.

Apple also designed the experience around selective escalation. Siri used its own systems for many requests, then offered ChatGPT when additional expertise appeared helpful.

That reduced the number of opportunities available to OpenAI. It also made the feature’s behavior less predictable for users.

A person opening the ChatGPT app knows which service will answer. A person speaking to Siri might receive an Apple response, a request for confirmation, or an offer to consult ChatGPT.

Uncertainty increases the effort required to form a habit. Users must remember which requests trigger the external model and how to ask for it directly.

ChatGPT Siri adoption also depended on Apple explaining the feature. The integration lacked the persistent visibility of a home-screen application or a dedicated navigation tab.

This matters because mobile distribution has several levels. Installation places software on a device. Prominent placement attracts attention. Default activation makes trial easier. Repeated value creates retention.

The Apple partnership delivered only parts of that chain. ChatGPT gained system access and brand association, but not default activation or ownership of the primary interface.

OpenAI’s filing frames those limitations as evidence against xAI’s foreclosure argument. A channel that barely adds users, OpenAI argues, cannot plausibly lock competitors out of a substantial market.

That legal argument is useful, but it also serves OpenAI’s interests. The company is presenting its weakest interpretation of the partnership while seeking summary judgment.

The redactions prevent outsiders from measuring the shortfall. We do not know the original forecast, the reduced forecast, or the final number of attributable users.

We also cannot separate product friction from other explanations. Apple Intelligence launched gradually across devices, regions, languages, and software versions. Those constraints narrowed the reachable audience.

Some consumers might have preferred the standalone ChatGPT app, which provided a more direct experience. Others might not have wanted generative AI inside Siri at all.

The court record supports a finding of underperformance against internal expectations. It does not establish one universal cause.

Still, the opt-in design offers the clearest visible mechanism. OpenAI expected distribution to create a funnel, but Apple constructed that funnel around user consent and platform control.

Those priorities reduced automatic exposure. They also left OpenAI dependent on Apple’s interface decisions, documentation, and product roadmap.

The Promised Halo Effect Met Apple’s Platform Control

The core reversal is simple: OpenAI gained coveted operating-system placement without gaining control over discovery, identity, or the user relationship.

When the companies negotiated, OpenAI sought a mutual two-year exclusivity period. Apple rejected that request.

The signed agreement explicitly described the arrangement as nonexclusive. It allowed Apple to integrate competing services or develop similar functionality itself.

Apple repeated that position publicly. During the 2024 launch, executives said the company intended to support additional AI models, including Google Gemini.

The distinction is central to the litigation. X and SpaceXAI alleged that the agreement blocked rivals and gave ChatGPT privileged access to potentially valuable iPhone prompts.

OpenAI says the contract contains no such restriction. Apple users could also download Grok, Gemini, Claude, or other applications through the App Store.

The filing adds a more damaging defense against the plaintiffs’ theory. Even if the placement favored ChatGPT, OpenAI argues that too few consumers used it to foreclose meaningful competition.

This produces an unusual courtroom posture. OpenAI is using its partnership’s disappointing performance as evidence that it lacked the power described by its rival.

The company’s argument does not erase the value of default platforms. Apple still controls important choices about presentation, permissions, routing, and access to system context.

Those choices determine whether an integration behaves like a destination, a background utility, or a rarely used fallback.

For OpenAI, the hoped-for halo effect depended on Apple lending more than technical access. It depended on Apple creating repeated moments where users clearly encountered ChatGPT.

The implementation gave Apple control over those moments. Apple decided how users enabled the service, when Siri suggested it, and what information required confirmation.

OpenAI supplied model capability but did not own the surrounding experience. That limited its ability to test onboarding, change prompts, promote accounts, or guide users toward repeat behavior.

This tension appears across platform partnerships. The platform owner wants outside capability without surrendering the customer relationship. The service provider wants distribution that builds its own identity.

Both companies can benefit when the arrangement works. Their incentives diverge when the provider wants deeper promotion or direct conversion.

Apple had another reason to preserve flexibility. Depending on one outside model would weaken its negotiating position and expose core features to another company’s roadmap.

A nonexclusive architecture allowed Apple to change providers or assign different models to different tasks. It also kept Apple positioned as the orchestrator of the experience.

That strategy became explicit on January 12, 2026. Apple and Google announced a multiyear collaboration for future Apple Foundation Models.

Their joint statement said the next generation would use Google’s Gemini models and cloud technology. The models would support future Apple Intelligence features, including a more personalized Siri.

The Google agreement does not prove that Apple has abandoned ChatGPT everywhere. It does confirm that OpenAI never secured an exclusive role inside Apple’s AI stack.

It also changes the meaning of the earlier partnership. ChatGPT increasingly looks like one component in Apple’s model portfolio, rather than the foundation of Apple Intelligence.

For Google, integration has different strategic value. Gemini already operates across Android, Search, Workspace, and other Google services. Supporting Apple’s models can strengthen its infrastructure position without requiring every interaction to become a Gemini account.

OpenAI depends more heavily on ChatGPT as a destination and customer relationship. Anonymous or Apple-mediated usage therefore provides less strategic value than direct engagement.

This comparison explains why identical placement can matter differently to two providers. The commercial value depends on who owns authentication, product discovery, telemetry, and the next user action.

The OpenAI Apple ChatGPT deal exposed that imbalance. OpenAI received reach, while Apple retained the controls that convert reach into routine behavior.

What the Filing Proves, and What It Leaves Hidden

The documents establish disappointing adoption, but they do not reveal enough data to grade the integration as a complete product failure.

OpenAI’s filing is advocacy, not an independent performance audit. Its purpose is to defeat antitrust claims before the case reaches trial.

The company benefits from showing that the agreement was nonexclusive and commercially weak. Both points undermine the allegation that ChatGPT’s placement substantially restricted competitors.

That does not make the disclosed evidence false. Court filings carry legal obligations, and OpenAI cites internal records, depositions, expert work, and the agreement itself.

However, readers should distinguish documented facts from the conclusions OpenAI draws from them.

The documented timeline is unusually clear. The agreement was signed in May 2024, announced in June, and launched in December. OpenAI lowered its user forecast by January 2025.

The company later described performance as dramatically and persistently below expectations. Several underlying figures and communications remain redacted.

Those redactions create major analytical limits. No public number shows how many people enabled ChatGPT, used it anonymously, connected an account, or became subscribers.

There is also no public retention curve. A low acquisition total can coexist with high satisfaction among the smaller group that adopted the feature.

The filing does not provide a device-level conversion rate. Without an eligible-device denominator, analysts cannot compare ChatGPT use across supported iPhones, iPads, and Macs.

It also does not separate Siri requests from Writing Tools or other entry points. Each surface solves different problems and carries different discovery costs.

OpenAI’s focus on incremental logged-in weekly active users might understate value delivered through anonymous interactions. Those interactions could still increase brand familiarity or improve Apple’s product experience.

Conversely, anonymous use might cost OpenAI computing resources without strengthening subscriptions or customer loyalty. The public record does not disclose that economic balance.

The definition of underperformance also depends on the original forecast. An ambitious estimate can make respectable usage look disappointing.

OpenAI and Apple had reason to project substantial reach in 2024. The partnership attracted worldwide attention and connected the best-known consumer chatbot with the iPhone.

Yet Apple Intelligence arrived through a staggered rollout. Hardware requirements excluded many existing devices, while regional and language availability expanded over time.

The standalone ChatGPT application was also established before the integration launched. Many interested users already had a direct route to OpenAI and needed little help from Siri.

That creates an attribution problem. If an existing ChatGPT user also activated the Apple extension, the partnership might increase convenience without creating an incremental user.

OpenAI specifically measured additional users it expected from the deal. Mature brand awareness made that harder than raw device reach suggested.

The legal context adds another complication. The filing characterizes the generative AI market as intensely competitive, citing Gemini, Claude, Meta AI, and Grok.

That competitive picture supports OpenAI’s defense. It also reflects the practical environment facing the integration.

Consumers did not need Siri to reach a capable model. They could install dedicated applications, use websites, or access AI through search engines and social platforms.

The best conclusion is therefore narrower than calling the partnership a total failure. Apple’s implementation did not deliver the incremental logged-in activity OpenAI expected.

That finding is significant because the placement looked exceptionally valuable. It shows that even elite distribution can disappoint when the service remains optional, indirect, and controlled by another interface.

It does not show that system integrations have no value. A deeper integration, clearer entry point, or better-defined task could produce different results.

It also does not resolve the antitrust case. The court must evaluate contractual language, market definition, competitive effects, standing, and the evidence offered by every party.

The public should treat OpenAI’s underperformance claim as a verified litigation position supported by cited records. The size and causes of that underperformance remain partly hidden.

Three Signals Will Show Whether Built-In AI Can Create Lasting Use

The next test is whether Apple can turn model choice into a clear user experience without recreating the friction that limited ChatGPT adoption.

The first signal is activation behavior. Apple can add model capabilities, but users must still discover, understand, and trust the route into them.

Any future disclosure about extension activation, account connections, repeat usage, or request volume would sharpen the picture. Rising opt-in and retention would weaken the idea that system AI cannot build habits.

Persistently low activation would strengthen the opposite conclusion. It would suggest that optional model access behaves more like a settings feature than a mass-market distribution channel.

The second signal is the role Google Gemini receives inside the next Siri. Apple and Google have described a deeper foundation-model collaboration than the original ChatGPT extension.

If Gemini operates underneath Apple’s interface without separate branding or activation, it might see far more use. That would show that integration depth matters more than model identity.

Such an outcome would not necessarily produce direct Gemini customers. It would instead support Google’s infrastructure and platform strategy while Apple retained the relationship.

If Apple exposes Gemini as another optional extension, the comparison will be cleaner. Similar friction and weak adoption would place more responsibility on Apple’s design.

Strong Gemini adoption under comparable conditions would pressure OpenAI’s product explanation. It would suggest that capability, task fit, or timing mattered alongside onboarding.

The third signal is OpenAI’s response to controlled distribution. The company can pursue integrations, strengthen its standalone application, or develop devices that give it greater ownership of the experience.

Each route addresses a different weakness. Partnerships offer reach, applications offer direct relationships, and dedicated hardware offers control over both interface and default behavior.

OpenAI does not need to abandon Apple to learn from the shortfall. It does need clearer measures for what an integration is supposed to achieve.

A partnership optimized for anonymous utility should be evaluated through request quality, satisfaction, and frequency. A customer-acquisition partnership needs visible login and retention paths.

Trying to achieve both through a hidden fallback creates confused incentives. Apple can claim better answers while OpenAI sees little measurable growth.

Developers and enterprise buyers should notice the same lesson. Adding an AI model to existing software does not guarantee adoption, even when the model and platform are household names.

A useful integration needs an obvious trigger, a repeatable task, understandable data controls, and feedback that confirms which system performed the work.

Knowledge workers face a related challenge. AI becomes valuable when it fits a recurring workflow and has access to the right context, not simply when another button becomes available.

Teams evaluating AI tools should therefore test active use rather than installation counts. They should measure successful tasks, repeat behavior, authenticated users, and what happens after the first answer.

The OpenAI Apple ChatGPT deal remains a major distribution experiment, even if its initial results disappointed OpenAI. Its most valuable output might be the lesson exposed through litigation.

Platform reach can open the door, but it cannot force users to walk through it. Product ownership, onboarding, and recurring utility still decide whether access becomes a habit.

Watch how Apple presents Gemini, whether ChatGPT activation changes, and how OpenAI builds direct user relationships. Those signals will show whether built-in AI becomes infrastructure or stays an overlooked option.