EU DMA Orders Google to Open Android to Claude & ChatGPT by 2027

By Deepak Bagada, CEO at SaaSNext & Principal AI Architect.

Last verified: August 2026 - European Commission DMA specification decisions of July 16, 2026, Article 6(7) and Article 6(11)

The OS Becomes the Gatekeeper

On July 16, 2026, the European Commission issued two binding specification decisions under the Digital Markets Act that force Alphabet's Google to open core parts of Android and Google Search to rival AI assistants - Claude, ChatGPT, Copilot, and any qualified competitor. This is the most aggressive interoperability order Brussels has aimed at Big Tech, and it targets the exact layer where the mobile agentic economy will be won or lost: the operating system.

The stakes are simple to state. Whoever controls the OS controls the assistant's reach - its wake word, its access to the screen, its ability to act inside other apps, its access to on-device AI hardware. Until now, only Google's own Gemini had full access to those hooks on Android. The Commission's finding is blunt: "competitors' AI assistants only have restricted access to key functionalities of the Google Android operating system. Without this access, alternative AI assistants are not competing on an equal footing." The orders change that by requiring Google to open 11 Android features to rival assistants by Android 18, no later than August 1, 2027, and to share anonymized search data with rivals and eligible AI companies from January 2027. Non-compliance carries fines of up to 10% of global annual turnover. The AI news desk has covered the specification proceedings since they opened in January, and this is the enforcement architecture that will now define mobile AI agents in Europe.

The 11 Features Google Must Open

The orders, issued under DMA Article 6(7), group the 11 Android features into five categories. Invocation: custom wake words, the long-press home gesture, and Circle to Search - the physical and voice triggers that currently default to Gemini. Context: on-device app data access, proactive suggestions based on screen content, and keyboard-integrated AI. Actions: autonomous control of other apps (booking, drafting, ordering) and access to OS settings. System services: the background hooks that let an assistant stay responsive. On-device ML: access to the neural processing units in modern phones, the specialized silicon that runs AI workloads locally.

Read as a whole, this is not a menu of app-store tweaks. It is a list of the capabilities that make an assistant feel native on a phone, and it is exactly the capability surface a mobile AI agent needs: wake up by voice, see the screen, act inside apps, run models on-device, and stay alive in the background. In practice, EU users will be able to set Claude or ChatGPT as their default assistant with full system integration, summon it by voice the way "Hey Google" works today, and have it perform actions inside other apps without leaving the conversation. The technical and policy template here - which hooks, what scope, what order - is worth studying even outside the EU, because it is the first concrete definition of what "OS-level agentic access" means.

Compliance Timeline

The search data order, under Article 6(11), is arguably the more consequential half. Google must share anonymized query, click, and ranking data with competing search engines and eligible AI chatbot providers, priced on fair, reasonable, and non-discriminatory (FRAND) terms. For AI assistants, this is the missing grounding signal: a general-purpose chatbot's answer quality is bounded by the freshness and breadth of the web signal it can see, and Google's index and click data are among the most valuable grounding signals in the world. OpenAI and Anthropic have built their own search capabilities largely without it; the January 2027 obligation changes that. Google retains the right to vet rivals for cybersecurity and data-protection risk, and only companies meeting the Commission's security and privacy criteria qualify.

Fines and the 10% Ceiling

The enforcement math changes the compliance decision. The DMA's penalty regime is revenue-linked: up to 10% of global annual turnover for non-compliance, with periodic penalty payments for ongoing violations. For Alphabet, whose turnover is on the order of $350 billion, 10% is a $35 billion order of magnitude. Compare that to the €4.1 billion fine Google paid in the earlier Android antitrust case: the DMA is designed to be structurally harder to negotiate around, because it targets the gatekeeper's architecture rather than a single abusive practice.

Google has publicly disagreed with the orders and warned that they "introduce unprecedented risks to user privacy, device security, and national security," and an appeal is likely. But under the DMA, an appeal delays implementation rather than eliminating the obligation; unless overturned, the specification decisions remain binding while the clock on the August 2027 deadline runs. The realistic compliance posture is build-now, litigate-in-parallel.

Compliance Architecture: What Android 18 Must Ship

For engineering teams, the interesting part is the technical shape of the fix. Google must expose assistant-interface capabilities at the platform layer rather than burying them inside Gemini. Concretely, that means a system-level assistant picker that is genuinely re-routable, an intent-routing layer that third-party assistants can register against, wake-word and gesture hooks that are not hard-coded to Gemini, a screen-context API with user consent, an app-control mechanism with scoped permissions, and NPU access via a vendor-neutral runtime. The pattern that fits this is familiar from mobile identity: declare capabilities, negotiate at first use, and route through a system mediator.

// Android 18 capabilities manifest: a rival assistant registers as a system assistant
{
  "assistant": "com.claude.agent",
  "bind": "android.service.assist.AssistService",
  "capabilities": {
    "invocation": ["wake_word", "long_press_home", "circle_to_search"],
    "context": ["screen", "keyboard", "proactive_suggestions"],
    "actions": ["app_control", "os_settings"],
    "system": ["background_runtime"],
    "ml": ["npu_inference"]
  },
  "consent": { "granular": true, "revocable": true }
}

// Screen-context access with per-app consent (DMA-compliant pattern)
val session = AssistantContextManager.requestSession(
    requested = setOf(ScreenContext, ActiveAppControl, NPUInference),
    scope = AppScope(userApprovedApps),
    reason = "Execute user request in current app"
)

The compliance architecture has three requirements that will survive any appeal. Interoperability must be symmetric: the same hooks Gemini uses, at the same latency, with no deliberate degradation. Consent must be granular and revocable: screen context and app control are sensitive, and the DMA template implies per-app, per-capability user approval. And the surface must be version-stable: an API that changes every release is interoperability in name only. The workflows section covers the agent-orchestration patterns these assistants will need once they can actually act inside the OS.

ROI: The Cost of Opening Versus the Cost of Not

Model the two paths. Compliance: opening 11 features across the Android stack is a real engineering program - platform APIs, permissioning, accessibility, NPU runtime, and the search-data pipeline - plausibly $250-$400 million in engineering and legal spend spread over 18 months, plus the strategic cost of letting competitors sit inside your OS. Non-compliance: a $35 billion worst-case fine, mandatory periodic payments, and the regulatory escalation that comes with ignoring a binding order. The arithmetic makes the decision for any rational operator: compliance is under 1.5% of the fine ceiling, and the fine is only the beginning - the Commission can impose behavioral remedies on top.

The search data order adds a different ROI dimension. Losing exclusive access to query-click-ranking data erodes the moat that has protected Google Search for two decades. But the FRAND pricing means Google is paid for the data, and the anonymization requirement preserves a meaningful distance between what rivals get and what Google itself uses. The realistic financial impact is a partial commoditization of search-derived signals rather than a collapse, which is why the engineering teams on both sides should be planning for a world where assistant answers are increasingly grounded in shared, priced, regulated data.

The OS-Level Agentic Access Debate

The order lands in the middle of a genuine architectural debate. One camp - largely the platforms - argues that deep OS integration is a security and privacy risk: an assistant with screen context, app control, and NPU access is a single prompt-injection incident away from acting on attacker instructions, and widening that surface multiplies the blast radius. Google's privacy and security warnings are not purely tactical; they are a real engineering position. The other camp - largely the assistant builders and the Commission - argues that an assistant that cannot see the screen or act in apps is not an agent at all, just a chat window, and that leaving the hooks exclusive to one vendor freezes the agentic economy in place.

Both positions have merit, which is why the technical details will be fought over for the next year. The safety answer that satisfies both sides is layered: system-level permissioning that grants capabilities per-app and per-action, on-device mediation that keeps sensitive context local, auditable consent that the user can revoke, and - critically - runtime security that watches what the assistant actually does with its access, the way enterprise agent protection watches tool calls in the cloud. The MCP directory is a useful map of the tool-interface layer these assistants will increasingly standardize on, and the same capability-access questions are being solved there in miniature.

What It Means

The practical consequences run in three directions. For assistant builders, Europe becomes the place where agentic access is a regulated right rather than a negotiated privilege, and the Android 18 surface is the first usable target. For Google, the orders convert the OS from a closed moat into a regulated utility, with the 10% turnover fine standing behind every compliance decision. And for everyone else, the definitional work - what features make an assistant real, what access is fair, what safety obligations travel with that access - is now being set in public by regulators, and it will leak well beyond the EU. Watch the latest AI news for the implementation fights; the architecture of mobile agentic access is being written now.