Brussels’ AI Omnibus Gives Frontier Labs More Time—and a Clearer Regulator

# Brussels’ AI Omnibus Gives Frontier Labs More Time—and a Clearer Regulator

*Lukas Hoffmann, Europe & Frontier Correspondent in Berlin — July 25, 2026*

The consequential Western AI development in the 24 hours from July 24 to July 25, 2026 did not come from a model laboratory. It came from Brussels.

The publication and entry into force on July 24 of Regulation (EU) 2026/1744↗, the Digital Omnibus on AI, changes the implementation path of Europe’s AI Act. It gives providers and deployers longer runways for high-risk systems, places oversight of systems based on general-purpose AI models more firmly with the European AI Office, expands support for smaller companies and makes two categories of abusive synthetic content explicitly prohibited.[[1]](digital-strategy.ec.europa.eu↗ [[2]](eur-lex.europa.eu↗ [[3]](consilium.europa.eu↗

That is the genuine in-window news. There was no supported new release during this period from OpenAI, Meta, Mistral, xAI, Cohere, AI21 Labs or Stability AI. Anthropic’s Claude Opus 5 announcement↗ also landed on July 24, but it is relevant here only as evidence of the continuing shift towards long-running agentic workloads—not as the subject of this analysis.[[4]](anthropic.com↗ [[5]](reuters.com↗ [[6]](anthropic.com↗

The Omnibus matters because that shift is changing what regulation must cover. Frontier models are no longer sold merely as systems that produce an answer. They increasingly sit inside agents that call tools, operate software, inspect documents, modify code and take approved actions. The economic unit is becoming the completed task; the regulatory unit remains a chain of providers, deployers, models and systems.

The Omnibus offers laboratories more implementation time, but it does not certify that their models—or the agents built around them—are safe.

What Brussels changed

The revised calendar is the most immediately commercial provision. Under the new framework, requirements for standalone high-risk AI systems apply from December 2, 2027. Requirements for high-risk systems embedded in regulated products apply from August 2, 2028.[[1]](digital-strategy.ec.europa.eu↗ [[7]](digital-strategy.ec.europa.eu↗ [[3]](consilium.europa.eu↗ [[8]](ec.europa.eu↗

The distinction is material. The first category includes systems operating in high-risk fields identified by the European framework, including biometrics, critical infrastructure, education, employment and migration. The second covers AI integrated into products such as machinery or toys.[[1]](digital-strategy.ec.europa.eu↗ [[7]](digital-strategy.ec.europa.eu↗ [[8]](ec.europa.eu↗ The Commission’s AI regulatory framework page↗ describes the wider implementation architecture, including guidance, standards work and support through the AI Act Service Desk.[[1]](digital-strategy.ec.europa.eu↗ [[9]](digital-strategy.ec.europa.eu↗

The Omnibus also makes several substantive governance changes:

• The European AI Office gains centralized oversight of AI systems based on general-purpose AI models, subject to specified roles and exceptions for national authorities, including in law enforcement and finance.[[1]](digital-strategy.ec.europa.eu↗ [[3]](consilium.europa.eu↗ [[10]](digital-strategy.ec.europa.eu↗
• Providers receive a three-month grace period for implementing transparency solutions for AI-generated content, ending on December 2, 2026.[[3]](consilium.europa.eu↗
• Simplified compliance modalities are extended beyond SMEs to small mid-cap companies, including reduced technical-documentation requirements, while access to regulatory sandboxes is broadened.[[1]](digital-strategy.ec.europa.eu↗ [[10]](digital-strategy.ec.europa.eu↗ [[8]](ec.europa.eu↗
• AI systems that generate non-consensual sexually explicit content or child sexual abuse material are explicitly prohibited, including systems associated with “nudification.”[[1]](digital-strategy.ec.europa.eu↗ [[11]](ai-act-service-desk.ec.europa.eu↗ [[3]](consilium.europa.eu↗ [[8]](ec.europa.eu↗
• Sandbox support is widened, with an EU-level sandbox expected to become operational by 2028, giving more developers a route to test systems under supervised real-world conditions.[[1]](digital-strategy.ec.europa.eu↗ [[10]](digital-strategy.ec.europa.eu↗ [[12]](euraxess.ec.europa.eu↗

The result is not simple deregulation. It is a trade: later application dates and lighter procedures for smaller businesses in exchange for a more legible supervisory structure and sharper prohibitions.

A delay tied to implementation readiness

The revised timing responds to a practical weakness in the original rollout: obligations are difficult to apply consistently when technical standards and compliance tools are incomplete. The Omnibus links the high-risk implementation path more closely to the availability of such support instruments.[[10]](digital-strategy.ec.europa.eu↗ [[9]](digital-strategy.ec.europa.eu↗

For laboratories, that can reduce the risk of engineering against requirements that later change. For customers, it can create time to classify systems, document workflows and decide where human review is necessary. Yet a longer runway can also prolong ambiguity if standards arrive late or enforcement practices differ across sectors and member states.

The tension is straightforward:

• More time can improve compliance quality if it is used to build evaluations, incident procedures and evidence trails around real deployments rather than to postpone them.
• More time can increase uncertainty if providers do not know which technical standards will ultimately be accepted as evidence of conformity.
• Centralization can improve consistency for general-purpose models, but only if the AI Office has sufficient expertise and enforcement capacity to examine systems that evolve faster than formal guidance.

The final regulation↗ establishes dates and responsibilities. It does not resolve every methodological question about how an agent should be evaluated once a general-purpose model is combined with tools, proprietary data and sector-specific instructions.[[2]](eur-lex.europa.eu↗ [[3]](consilium.europa.eu↗

Agents make token economics a governance issue

Three days before the Omnibus publication, Google announced Gemini 3.6 Flash, Gemini 3.5 Flash-Lite and Gemini 3.5 Flash Cyber↗. That July 21 release is outside the news window, but it illustrates the market the regulation must address.[[13]](blog.google↗

Google priced 3.6 Flash at $1.50 per million input tokens and $7.50 per million output tokens, while saying it used 17% fewer output tokens than Gemini 3.5 Flash on the Artificial Analysis Index. Flash-Lite was priced at $0.30 per million input tokens and $2.50 per million output tokens, with a reported throughput of 350 output tokens per second.[[13]](blog.google↗

Those figures matter, but headline token prices do not equal the cost of an agentic task. An agent may reason through several steps, call tools, inspect results, revise a plan, invoke subagents and repeat failed actions. The bill depends not only on the price of each token but also on output length, reasoning steps, tool calls and execution loops.

Google says 3.6 Flash takes fewer reasoning steps and tool calls in multi-step workflows, and produces fewer unwanted code edits and execution loops. If that holds in a customer’s own environment, it can reduce both inference expenditure and operational risk: fewer actions mean fewer opportunities for an agent to make an erroneous change.[[13]](blog.google↗ But the published figures combine vendor claims with named benchmark results. They are not substitutes for testing on the actual permissions, interfaces and failure conditions of a regulated deployment.

Computer use changes the risk boundary

Both 3.6 Flash and Flash-Lite offer computer use as a built-in client-side tool through Google’s interfaces, enabling agents to work across browser and desktop environments.[[13]](blog.google↗ OpenAI’s separately dated Presence platform↗, introduced on July 22, similarly shows the move towards managed agents that interact with company systems and perform approved actions.[[14]](openai.com↗

Presence allows enterprises to define policies, guardrails and escalation rules, and to transfer cases to a human when an agent reaches a scenario outside its boundaries. It also provides simulation and evaluation tools for common requests, edge cases and high-risk scenarios before production deployment.[[14]](openai.com↗

These controls are more informative than a general claim that a model is “safe.” In an agentic system, risk depends on what the model can access, what actions it can execute, when it must stop and what evidence remains after an action. A capable model with constrained permissions may present less operational exposure than a weaker model with broad credentials and no reliable escalation path.

Model evaluation measures capability under defined conditions. Deployment governance determines what that capability can touch. Neither should be mistaken for the other.

This is why the AI Office’s centralized role around general-purpose AI is important. A single model can support many downstream systems, but each deployment may produce a different risk profile. The Office can provide a common supervisory point for the model layer; sectoral and national authorities still have to address the environments in which systems operate.[[1]](digital-strategy.ec.europa.eu↗ [[3]](consilium.europa.eu↗

Cyber pilots show how capability control may work

Cybersecurity makes the dual-use problem particularly visible. Google’s July 21 announcement says Gemini 3.5 Flash Cyber is fine-tuned to find and fix vulnerabilities and is deployed through CodeMender, where multiple agents work together to generate a combined report. Because of the technology’s dual-use nature, Google said it would be available only to governments and trusted partners through a limited-access pilot.[[13]](blog.google↗

That restriction is not proof that misuse is impossible. It is an access-control decision around a system whose defensive utility and offensive potential cannot be cleanly separated. Google also says 3.6 Flash includes enhanced safeguards for cyber-offence and chemical, biological, radiological and nuclear misuse, with greater resistance to jailbreaks while attempting to minimise refusals of beneficial requests.[[13]](blog.google↗

The Omnibus and this pilot model are aligned in three practical respects:

• Both place weight on identifiable providers and controlled access, rather than assuming that a model-level refusal alone will govern downstream conduct.
• Both imply a need for evaluation beyond generic benchmarks, because cyber performance depends on orchestration, tools, target environments and the ability to validate or execute a proposed change.
• Both support centralized scrutiny of general-purpose capabilities, while leaving room for restrictions around specialised applications and particular classes of users.

The caveat is equally important. The Google release↗ reports competitive performance on CyberGym and describes stronger frontier safeguards, but a benchmark cannot establish how a system will behave against novel targets, under adversarial prompting or when several agents operate together.[[13]](blog.google↗ Regulatory oversight therefore needs access to procedures, deployment controls and evidence—not merely a score.

Different incentives for hyperscalers and European challengers

US hyperscalers enter this implementation period with advantages that the Omnibus does not erase. They can spread compliance engineering across model platforms, enterprise products and cloud infrastructure. They can also offer managed deployment layers that package evaluations, policy controls and escalation mechanisms alongside inference.

The July 22 OpenAI Presence announcement↗ is a clear example: deployments are managed through OpenAI’s forward-deployed engineers and selected systems integrators, rather than offered as a purely self-service product.[[14]](openai.com↗ [[15]](openai.com↗ That model gives the provider more visibility into deployment conditions, while potentially making customers more dependent on its governance stack.

European challengers such as Mistral face a different calculation. Mistral made no new announcement in the July 24–25 window, but its existing emphasis on enterprise workflows, self-hosting, data control and European infrastructure provides relevant competitive context.[[16]](mistral.ai↗ [[17]](mistral.ai↗ [[18]](mistral.ai↗ [[19]](mistral.ai↗ The extension of simplified requirements to small mid-caps and the expansion of sandboxes may disproportionately help European companies that cannot maintain hyperscaler-sized compliance teams.[[1]](digital-strategy.ec.europa.eu↗ [[10]](digital-strategy.ec.europa.eu↗ [[8]](ec.europa.eu↗

Yet lighter documentation does not remove the fixed cost of proving that a system behaves appropriately. A smaller provider still needs evaluation methods, technical records and incident processes that enterprise buyers will trust. The strategic opportunity is therefore not merely “European compliance.” It is to offer deployments whose data location, permissions, auditability and operating boundaries are understandable to regulated customers.

The questions Brussels has postponed, not answered

The Digital Omnibus↗ provides a more workable schedule, but its success will depend on implementation.[[2]](eur-lex.europa.eu↗ Four questions now deserve close attention.

• Will harmonised standards arrive early enough for providers to design stable compliance programmes before December 2027 and August 2028?
• Can the AI Office develop sufficient technical and investigative capacity to supervise general-purpose models and the systems built on them without creating a procedural bottleneck?
• Will the extra time produce better evidence, including deployment-specific agent evaluations, or simply defer difficult classification and accountability decisions?
• Can transparency solutions due by December 2, 2026 remain effective as generated content and agentic workflows change, rather than becoming a one-off documentation exercise?

The Commission’s regulatory framework↗ offers support mechanisms, while the Omnibus sharpens the calendar and institutional division of labour.[[1]](digital-strategy.ec.europa.eu↗ [[9]](digital-strategy.ec.europa.eu↗ What neither can do is turn a regulatory deadline into empirical proof of reliability.

For Western frontier laboratories, the message from Brussels is consequently more exacting than the language of “delay” suggests. They have gained time. They have also been given a clearer indication of where scrutiny will concentrate: general-purpose model oversight, agentic deployment controls, transparency, specialised dual-use access and evidence that survives beyond the benchmark sheet.

That is breathing room, not absolution.

---

References

1. <digital-strategy.ec.europa.eu↗> 2. <eur-lex.europa.eu↗> 3. <consilium.europa.eu↗> 4. <anthropic.com↗> 5. <reuters.com↗> 6. <anthropic.com↗> 7. <digital-strategy.ec.europa.eu↗> 8. <ec.europa.eu↗> 9. <digital-strategy.ec.europa.eu↗> 10. <digital-strategy.ec.europa.eu↗> 11. <ai-act-service-desk.ec.europa.eu↗> 12. <euraxess.ec.europa.eu↗> 13. <blog.google↗> 14. <openai.com↗> 15. <openai.com↗> 16. <mistral.ai↗> 17. <mistral.ai↗> 18. <mistral.ai↗> 19. <mistral.ai↗>