Tokens got cheaper. Permission got expensive.

Takeaway

Contents

What is in this edition

01 · The structural read

Cheap tokens, expensive permission

For three years the enterprise AI business case has been written against one variable: the cost of inference. That variable kept cooperating. It cooperated again on September 1, when Anthropic made Claude Fable 5.1 generally available at the same list price as its predecessor, dropped cache-read pricing to $0.25 per million tokens, and reported that some highly agentic workloads now cost up to 45 percent less to run CITED C01. Google shipped Gemini 3.8 Flash and a Gemini 3.8 Flash Cyber variant on September 2 CITED C05. Tencent, Z.ai and Alibaba's Qwen team all put new flash-tier models into the market inside the last eight days CITED C06. Per unit of work, the model layer is getting cheaper on a weekly cadence.

What is not getting cheaper is permission. Three items landed in the same 48 hours, from three unrelated directions, and each one raises the cost of the conditions under which an agent is allowed to operate.

The grid. Texas became the first major data-center market to freeze new large-load grid connections while it audits its own queue. Interconnection requests from data centers and other large loads have gone from roughly 48 GW in 2023 to more than 474 GW, a figure grid operators themselves describe as substantially speculative, and Pennsylvania has since moved in the same direction VERIFIED C08. Bloomberg's August 12 analysis of the wider US picture put the share of requested data-center load that utilities are likely to actually commit to at about 28 percent of 1,066 GW CITED C09.

The agent supply chain. Palo Alto Networks acquired Console, an agentic IT-operations platform, in a deal reported at $500 million CITED C11. A startup called AIR emerged with $50 million specifically to discover agents running inside enterprises and inspect the plug-ins, skills and MCP servers they load, claiming that about 27 percent of the public add-ons it evaluates fail its security criteria FLAG C10. In the same window, attackers were actively exploiting CVE-2026-0768 in Langflow, an open-source agent orchestration framework, to execute code and steal cloud and model API credentials CITED C12.

The capability threshold. OpenAI said its forthcoming Astra model has crossed the Critical bar for cyber capability under its own Preparedness Framework, which triggers additional controls on access, monitoring and deployment; company material cited two previously unknown vulnerabilities found during evaluation CITED C03. Anthropic already draws the same line differently, releasing Fable 5.1 broadly while restricting the more capable Mythos 5.1 to vetted organizations CITED C01. Reuters reported on September 2 that OpenAI told two House Democrats its engineers are developing automated shutdown capabilities for AI systems CITED C04.

02 · Evidence discipline

Shipped, announced, and still a plan

Same discipline as every edition. An item is shipped if a buyer can use it today. It is announced if it is dated and specified but not yet generally available. It is a plan if it is a target, a pilot, a memorandum, or a forecast. Treat only the first column as procurable.

03 · Frontier ledger

The frontier ledger, read globally

We read the frontier by jurisdiction, because jurisdiction is a procurement axis in regulated sectors. Where a model may be hosted, under what license, and which capability tier is sold into which market are now three separate questions with three separate answers.

04 · Financial services

Financial services: the rulebook is deliberately blank

The win. Banks that moved first on agentic AI did it in the least glamorous places and are still there: financial-crime detection, regulatory-change triage, controls testing, and continuous transaction monitoring. These are procedural, evidence-producing tasks where every agent step is traceable and a human approves the output. That is the deployment pattern with the most supervisory tolerance today CITED C19.

The constraint. In April 2026 the Federal Reserve, OCC and FDIC issued revised interagency model risk management guidance, and it says explicitly that generative and agentic AI are novel and rapidly evolving and are not within its scope, with a request for information signaled to follow CITED C20. Meanwhile FINRA moved agentic AI from emerging technology to an active supervisory priority in its 2026 Annual Regulatory Oversight Report and treats it as a distinct supervisory risk category CITED C19. Read those two together. Examiners will ask about agent governance. The prudential guidance you would normally map your answer to has told you it does not cover this. The absence of a rule is not a grace period; it is an instruction to show your own work.

The control. Build an agent inventory that carries, per agent: the business process it touches, the systems and credentials it can reach, the human approval gate and who holds it, the retention location for its decision log, and the named owner. Then rehearse the examiner conversation against it. If your answer to how do you know this agent did not exceed its mandate depends on a vendor dashboard you do not control, that is the gap to close first.

05 · Healthcare

Healthcare: the comment window is the control point

The win. The FDA has authorized more than 1,000 AI and machine-learning enabled devices, and clearances have been running at record annual volume. The pathway works for narrow, well-characterized tools CITED C13.

The constraint. Generative AI is a different regulatory object, and the agency has said so. On August 18, 2026 the FDA published a discussion paper on how it intends to regulate generative-AI-enabled medical devices and opened a public comment window on docket FDA-2026-N-7874 that runs to October 19, 2026 VERIFIED C13. Separately, Mosaic Clinical Technologies, a unit of Radiology Partners, filed a citizen petition the same day asking the agency to clarify when commercially distributed vision-language models used in diagnostic imaging must be regulated as devices CITED C21. The boundary between clinical decision support and a regulated device is being drawn right now, in public, by whoever shows up.

The other constraint. Aesto Health, a vendor that migrates and archives patient records during EHR transitions and acquisitions, disclosed a breach affecting more than 9.5 million people. The intrusion into part of its AWS environment occurred in December 2025; the scope of affected protected health information was not confirmed until May 2026 CITED C22. That is a seventeen-month gap between compromise and public disclosure, in a vendor most patients have never heard of. Your AI vendor inventory and your records-custody inventory are the same inventory, and most health systems maintain them separately.

The control. Two actions this month. First, decide whether you are filing on FDA-2026-N-7874, and if you are, file on the boundary question specifically, because that is what will determine whether your ambient documentation or imaging triage tool is a device. Second, run a custody map: for every AI or data vendor, where does patient data physically rest, under whose keys, and what is the contractual breach-notification clock. Anthropic's Enterprise Frontier Safeguards approach, customer-held logs in the customer's own S3, Azure Blob or Google Cloud Storage under the customer's own keys, is worth using as the benchmark you hold other vendors to VERIFIED C02.

06 · Manufacturing

Manufacturing: operating hours are the only humanoid metric that clears diligence

The win. Two programs have numbers a diligence team can use. Figure's deployment at BMW Spartanburg runs ten-hour shifts loading sheet-metal parts into welding fixtures, with more than 90,000 parts handled and placement accuracy reported above 99 percent. Agility Robotics reports more than 65,000 accumulated operating hours across nine commercial facilities, and states annual production capacity of 10,000 Digit units at its RoboFab plant CITED C16. Those are company-reported figures, but they are the right kind of figure: hours, units, error rate, named site.

The constraint. Tesla's Optimus program is estimated at roughly 1,000 to 1,200 units across Fremont and Giga Texas as of mid-2026, with the Fremont Model S and X line converted to Optimus assembly, but no external sales and no published uptime figures FLAG C16. More broadly, independent analysis in July 2026 concluded that no humanoid from any manufacturer has been deployed above the low hundreds of units in a sustained commercial environment CITED C23. The category is real and the category is small. Both things are true.

The control. Write your humanoid or mobile-manipulation RFP so that the first scored section is operating evidence, not capability. Require: named reference site, cumulative operating hours, mean time between interventions, task error rate with a stated denominator, and who performs field maintenance. Score announcements at zero. This one change will collapse a twelve-vendor long list to three inside a week, and it is defensible to your board because it mirrors how you already buy any other piece of line equipment.

07 · Energy

Energy: your AI roadmap now has an interconnection dependency

The win. AI load is the first thing in two decades to restart real growth planning on the US grid. Data-center electricity demand has moved from roughly 23 GW in 2023 to about 42 GW in 2026, and AI racks draw 50 to 100 kW against 5 to 10 kW for conventional racks. That is a genuine industrial-capacity story with a long tail of transmission, generation and equipment work attached CITED C09.

The constraint. The queue is not the pipeline. Requests to connect to the Texas grid from data centers and other large loads have risen from roughly 48 GW in 2023 to more than 474 GW, and on the strength of that Governor Abbott directed an audit of every data center in the ERCOT interconnection queue on August 3, 2026, including whether projects rely on taxpayer-funded incentives, their water use, and their on-site generation plans. Texas then became the first major hub to freeze new connections while it works through that audit, and Pennsylvania followed with a similar move VERIFIED C08. Nationally, Bloomberg's analysis put likely utility commitments at about 28 percent of the 1,066 GW requested CITED C09. The industry term for the rest is ghost demand: speculative queue positions taken cheaply, without a firm customer or funding behind them.

The control. If any part of your AI program depends on capacity landing at a specific site on a specific date, that date is now an assumption to be tested, not a plan. Ask your cloud or colocation provider three questions in writing: which specific substation and queue position serves the capacity you have been promised, what is its current energization date, and what is the contractual remedy if it slips. Then build the fallback: multi-region placement, an inference-tier downgrade path, and a workload-deferral policy that names which agents get throttled first when capacity is short. Regulated firms with real-time obligations should treat this the same way they treat any other single-point utility dependency.

08 · Cross-sector risk

The agent supply chain became an attack surface

Three items from the same 48 hours describe one problem. Attackers are actively exploiting CVE-2026-0768 in Langflow, an open-source framework used to build agent workflows, to run unauthenticated remote code and harvest credentials, including model API keys and cloud credentials stored in exposed environments CITED C12. AIR launched with $50 million to discover agents inside enterprises and inspect the plug-ins, skills and MCP servers they load, alongside a category that already includes Noma Security, Zenity and Astrix FLAG Source C10. And Palo Alto Networks bought Console to put agents inside its own Cortex security operations stack CITED C11.

The pattern is that agent orchestration frameworks sit between models, databases, APIs and cloud infrastructure. Compromising one yields something more valuable than the application: the credentials that open everything behind it. Meanwhile the same agents are being granted operator privileges inside security operations. Both directions of that trade are moving faster than most enterprise change-control processes.

Three controls that cost little and pay this quarter.

1. Inventory the connectors, not just the agents. For every agent in production, list every plug-in, skill and MCP server it can load, and who approved each one. If you cannot produce that list in an hour, that is the finding.
2. Scope credentials to the task, with expiry. Agent credentials should be short-lived and narrowly scoped. The Langflow incidents are credential-theft incidents; a long-lived broad-scope key turns a framework bug into a cloud breach.
3. Patch agent frameworks on the internet-facing schedule. Langflow, and the rest of the orchestration layer, should sit in the same patch tier as your public web servers and identity infrastructure, not in the experimental tier where most of it currently lives.

09 · Governance clock

The governance clock, read honestly

Market commentary continues to assert that EU AI Act high-risk obligations for financial services became applicable on August 2, 2026. That is not correct, and it matters because firms are budgeting against it. Here is the primary-source position.

The planning consequence of the deferral is not relief. It is that the compliance work moved from a date you could point at to a date far enough out that it will lose an internal budget fight this autumn, and then arrive with eighteen months of accumulated system debt attached. Firms that keep building the technical file through 2027 will be the ones who are not doing it under time pressure in 2028.

10 · Practitioner section

Build the agent permission ledger

Every edition ends with something you can build. This week it is a single artifact that answers the three permission questions raised above in one place. It is a table, it fits on a page, and it is the fastest way we know to turn this week's news into a defensible position.

One row per production agent. Six columns.

Two rules make it useful rather than decorative. First, the ledger is owned by the business, not by IT, because five of the six columns are business decisions. Second, an agent that cannot be described in a complete row does not run in production. That rule is unpopular for about a month, and then it becomes the reason your examination goes well.

11 · Field notes

Five things worth knowing this week

1. A cache-read price cut is a bigger deal than a benchmark. Long-running agents re-read context constantly. Cache-read pricing dropping to $0.25 per million tokens changes the unit economics of exactly the workloads regulated firms run: document review, monitoring, reconciliation CITED C01.
2. Customer-held logs is now a benchmark you can hold every vendor to. Anthropic's Enterprise Frontier Safeguards keeps activity data in the customer's own S3, Azure Blob or Google Cloud Storage under the customer's keys, while automated misuse monitoring still runs, with no human review by the vendor. Whether or not you buy it, ask every other AI vendor to match it in writing VERIFIED C02.
3. Ghost demand is a governance failure, not a forecasting failure. Queue positions were cheap, so everyone took several. The fix Texas chose, an audit of who is real, is the same fix your agent estate needs VERIFIED C08.
4. Broadcom's quarter contains both halves of the story. AI semiconductor revenue up 221 percent year on year to $16.7 billion, and a Q4 guide below consensus, in the same company-reported release CITED C17.
5. Open weights from China are now a residency option, not just a price option. GLM-5.3-Flash under MIT terms with a 1M context window, and an open-weight Qwen4-architecture preview, both landed on August 26. For an air-gapped or in-country deployment that changes the shortlist. It also imports a supply-chain review you did not previously have to run CITED C06.

12 · Limits

What we are not asserting

• We are not asserting that Fable 5.1, Gemini 3.8 Flash, Grok 4.6, Hy4 or GLM-5.3-Flash outperform one another. All published benchmark figures referenced here are vendor-stated and are not independently reproduced by us.
• We are not asserting that the AIR figure of 27 percent failing add-ons is representative. It is a single vendor's screen of its own corpus, disclosed in a funding announcement. It is carried as FLAG for that reason.
• We are not asserting that Astra is more or less dangerous than any other frontier model. We are reporting that OpenAI applied its own Critical designation and that this triggers additional deployment controls.
• We are not asserting that the Texas freeze is permanent or that 474 GW represents real demand. The point of the audit is that nobody currently knows which portion is real.
• We are not asserting a causal link between Broadcom's forward guidance and grid constraints. They are reported adjacently because both belong in a capacity plan.
• Humanoid figures for Figure, Agility and Tesla are company-reported or third-party estimates, not audited. We use them to rank evidence quality, not to rank vendors.

13 · Sources

Research base for this edition

Every quantitative claim above carries an ID. Each ID resolves here. Where market commentary conflicted with primary legal sources, notably on EU high-risk applicability, we followed the primary legal sources and said so in section 09.

• C01 Anthropic Claude Fable 5.1 and Mythos 5.1, generally available September 1, 2026. List price unchanged at $10 per million input and $50 per million output tokens; cache reads reduced to $0.25 per million; vendor states up to 45 percent lower cost on some highly agentic workloads; self-reported Terminal-Bench-Science 52.6 against 24.7 for Fable 5. Mythos 5.1 restricted to vetted organizations. Benchmark figures are vendor-stated. LLM Stats, Claude Fable 5.1 launch analysis, September 1, 2026 · Tech Startups daily roundup, September 2, 2026
• C02 Anthropic Enterprise Frontier Safeguards, announced September 1, 2026. Activity data stored in customer-controlled cloud infrastructure (Amazon S3, Azure Blob Storage, Google Cloud Storage) under customer encryption keys and access policies; automated misuse monitoring retained without human review by Anthropic; no charge stated; phased rollout beginning later in autumn 2026; interim zero data retention on Fable 5 and Fable 5.1 for eligible customers; developed with more than 100 customers across financial services, healthcare, manufacturing, telecom, law, retail and the public sector, with AWS, Google Cloud and Microsoft Azure as cloud partners. Anthropic, Developing Enterprise Frontier Safeguards with our customers · CNBC, September 1, 2026
• C03 OpenAI Astra cyber capability classification. OpenAI states Astra has reached the Critical threshold for cyber capability under its Preparedness Framework, triggering additional controls on access, monitoring, model behavior and deployment; two previously unknown vulnerabilities identified during evaluation. Astra is not generally available. Tech Startups, September 2, 2026, citing The Information and OpenAI · LLM Stats news wire, September 2, 2026
• C04 OpenAI automated shutdown capabilities. Reuters reported on September 2, 2026 that OpenAI told two House Democrats in a letter that its engineers are developing automated shutdown capabilities for AI systems. Described as under development, not deployed. LLM Stats news wire aggregating Reuters via Techmeme, September 2, 2026
• C05 Google model and product releases. Gemini 3.8 Flash and Gemini 3.8 Flash Cyber listed as released September 2, 2026. Agentic video understanding added to Gemini, allowing the model to select what to watch and at what speed. Gemini-powered remembered-items capability shipped into Android Find Hub in the September 2026 feature drop. LLM Stats, model release log · Tech Startups, September 2, 2026, citing The Verge
• C06 China and open-weight releases, August 25 to September 2, 2026. Tencent Hy4 preview (August 28). Z.ai GLM-5.3-Flash, first natively multimodal GLM-5, 320B total and 18B active, 1M context, MIT weights, self-reported DeepSWE 63.4 against GLM-5.2 at 46.2 (August 26). Qwen3.8-Flash and Qwen3.8-Flash-Next, an open-weight preview of the Qwen4 architecture, 6B active, 262K native context (August 26). IBM Granite 4.2 30B and 3B (August 25). Tencent-backed accelerator maker Enflame Technology reported online-offering oversubscription above 6,000 times on a Shanghai listing raising roughly 6.1 billion yuan, about $900 million. All benchmark figures vendor-stated. LLM Stats, GLM-5.3-Flash launch · LLM Stats, Qwen3.8-Flash-Next launch · Tech Startups on Enflame, September 2, 2026, citing Reuters
• C07 xAI product position. Grok 4.6 launched August 12, 2026. Grok 4.5, released July 2026, positioned by xAI as its strongest model for software engineering and long-running computer tasks, and placed fourth on one public composite leaderboard at 93.0 percent. Grok Bot entered early beta on August 11, 2026, allowing a named bot to be assigned persistent tasks and to continue working while the user is offline; business and enterprise workspaces with licenses and organization controls are available. Leaderboard position is a third-party composite, not an audited result. xAI newsroom · LLM Stats composite leaderboard
• C08 Texas large-load interconnection freeze. Texas became the first major data-center hub to freeze new grid connections for data centers pending an audit. Requests from data centers and other large loads to connect to the ERCOT grid rose from about 48 GW in 2023 to more than 474 GW. Governor Greg Abbott directed an audit of all data centers in the ERCOT interconnection queue on August 3, 2026, covering reliance on taxpayer-funded incentives, water use and on-site generation plans, to be implemented by the Public Utility Commission of Texas and ERCOT. Pennsylvania subsequently made a similar move. Freeze reported September 1, 2026. Utility Dive, Texas hits pause on data center interconnections · BNN Bloomberg, September 1, 2026
• C09 US data-center load and grid commitment. Bloomberg analysis published August 12, 2026 finds grid operators and utilities likely to commit to about 28 percent of the 1,066 GW requested for data-center projects, with the remainder attributed to speculative or duplicated requests. US data-center electricity demand rose from about 23 GW in 2023 to about 42 GW in 2026. AI racks draw 50 to 100 kW against 5 to 10 kW for conventional racks. Bloomberg, August 12, 2026 · MarketScale, US grid asset rethink
• C10 Agent software supply chain security. AIR emerged from stealth on September 2, 2026 with $50 million raised across two seed rounds led by Sequoia and Greenoaks, founded by Yair Saban and Niv Hoffman. Platform discovers AI agents inside enterprises and inspects the plug-ins, skills and MCP servers they use. The company states that roughly 27 percent of publicly available add-ons and skills it evaluates fail its security criteria. This is a vendor-stated figure about its own corpus and is carried as FLAG pending independent replication. Named competitors include Noma Security, Zenity and Astrix. Tech Startups, September 2, 2026, citing TechCrunch
• C11 Palo Alto Networks acquisition of Console. Palo Alto Networks acquired Console, an AI-native agentic workflow platform, with plans to integrate it into the Cortex platform for alert analysis, incident prioritization and action in customer environments. Terms were not disclosed by the company; TechCrunch reported a figure of $500 million citing sources. LLM Stats news wire aggregating TechCrunch, September 2, 2026 · SecurityWeek
• C12 Langflow exploitation. CVE-2026-0768 in Langflow, an open-source framework for building AI applications and agent workflows, permits unauthenticated remote Python code execution. Active exploitation observed, with attackers extracting credentials including OpenAI API keys and AWS credentials from exposed environments. Tech Startups, September 2, 2026, citing BleepingComputer
• C13 FDA generative AI medical device regulation. The FDA published a discussion paper on August 18, 2026 outlining its intended regulatory approach for generative-AI-enabled medical devices and opened a public comment window under docket FDA-2026-N-7874 on Regulations.gov, closing October 19, 2026. The agency has authorized more than 1,000 AI and machine-learning enabled devices cumulatively. FDA press announcement · FDA AI medical device list, statistics and trends
• C14 Cursor Cloud Agents. In August 2026 Cursor removed the requirement for a connected GitHub or other source-control provider to start a Cloud Agent, allowing work to be saved to a Cursor Origin repo, previewed in browser and published to Vercel with a live URL. Cursor offers team and enterprise plans with administrative controls, usage analytics, single sign-on, model controls and compliance features. Cursor release notes, August 2026 · Cursor product overview
• C15 Meta model and device actions. Meta released Muse Voice Transcribe via API, described by the company as state of the art in streaming speech to text and trained on more than 70 languages. Meta disabled camera functionality on thousands of pairs of smart glasses after detecting tampering with the recording indicator light, stating that fewer than one tenth of one percent of glasses sold had been altered. Performance claim is vendor-stated. Tech Startups, September 2, 2026, citing Semafor · LLM Stats news wire
• C16 Humanoid robotics deployment evidence. Figure AI's BMW Spartanburg deployment runs ten-hour shifts loading sheet-metal parts into welding fixtures, with more than 90,000 parts handled and placement accuracy reported above 99 percent. Agility Robotics reports more than 65,000 accumulated operating hours across nine commercial facilities, with stated RoboFab production capacity of 10,000 Digit units annually. Tesla's Optimus program is estimated at 1,000 to 1,200 units across Fremont and Giga Texas as of mid-2026, with the Fremont Model S and X line converted to Optimus assembly, no external sales and no published uptime figures. Tesla stated on its August 5, 2026 Q2 earnings call that production was starting soon. Figures are company-reported or third-party estimates, not audited; the Optimus deployment estimate is carried as FLAG. Tesla Optimus and Agility Digit comparison, 2026 · The AI Insider, state of humanoid robotics, August 21, 2026
• C17 Broadcom fiscal Q3 2026 results, company-reported, September 2, 2026. Revenue up 86 percent year on year to $29.59 billion against consensus of $29.36 billion. AI semiconductor revenue up 221 percent to $16.7 billion. Q4 revenue guidance below consensus. Separately, Nvidia's chief executive urged G20 ministers on September 2, 2026 to build more data centers. LLM Stats news wire aggregating CNBC via Techmeme, September 2, 2026 · NVIDIA newsroom, agent development platform
• C18 European AI infrastructure. Spanish semiconductor startup iPronics raised $125 million including participation from Nvidia, for programmable photonic chips that move data between computing systems using light rather than electrical signaling, targeting the AI data-center networking bottleneck. Mistral continues a regional-inference and European-compute programme. Tech Startups, September 2, 2026, citing Reuters · Mistral AI, regional inference and European compute
• C19 FINRA and banking agentic AI supervision. FINRA's 2026 Annual Regulatory Oversight Report moves agentic AI from emerging technology to an active supervisory priority and classifies it as a distinct supervisory risk category. Current bank deployments concentrate on procedural, auditable work: financial-crime detection, regulatory-change triage, controls testing and continuous transaction monitoring, in governed environments where agent decisions are traceable and a human approves outputs. FINRA 2026 agentic AI classification, analysis for banks and lenders · Agentic AI and compliance in 2026
• C20 US interagency model risk management guidance. In April 2026 the Federal Reserve, OCC and FDIC issued revised interagency model risk management guidance, SR 26-2, which states that generative and agentic AI are novel and rapidly evolving and are not within its scope. The agencies signaled plans to issue a request for information addressing banks' use of AI. CCG Catalyst commentary on SR 26-2 scope · AI governance in banking, 2026 guide
• C21 Diagnostic imaging vision-language model petition. Mosaic Clinical Technologies, a unit of Radiology Partners, filed a citizen petition on August 18, 2026 asking the FDA to clarify when commercially distributed AI vision-language models used for diagnostic imaging must be regulated as medical devices. FDA generative-AI medical device brief, 2026 · FDA press announcements
• C22 Aesto Health breach. Aesto Health, a provider of software used by medical organizations to migrate, archive and access patient information during electronic health record transitions and acquisitions, disclosed a cyberattack affecting more than 9.5 million individuals. An unauthorized actor accessed part of its Amazon Web Services environment in December 2025; the scope of potentially compromised protected health information was not confirmed until May 2026. Tech Startups, September 2, 2026, citing BleepingComputer
• C23 Humanoid deployment scale, independent analysis. A July 18, 2026 analysis concluded that no humanoid robot from any manufacturer has been deployed above the low hundreds of units in a sustained commercial environment as of mid-2026. Technology.org, humanoid robots in 2026, what is actually deployed
• C24 EU AI Act and AI Digital Omnibus, primary legal position. The Council of the EU gave final approval to the AI Digital Omnibus on June 29, 2026, following European Parliament endorsement on June 16, 2026 by 423 votes to 57 with 174 abstentions. The Omnibus was published in the Official Journal on July 24, 2026 and entered into force on July 27, 2026, six days before the AI Act's original August 2, 2026 high-risk deadline. High-risk obligations for standalone Annex III systems are deferred by 16 months to December 2, 2027; for AI embedded in Annex I regulated products, by 12 months to August 2, 2028. Article 50 transparency obligations applied from August 2, 2026 and were not deferred. Where market commentary asserted that financial-sector high-risk obligations applied from August 2, 2026, we follow the primary sources above. Council of the EU, final green light, June 29, 2026 · Gibson Dunn, postponed high-risk deadlines