Skip to content
FAQ
AI & ML Analysis

AI Week, Aug 31–Sep 6, 2026: EU Puts ChatGPT in Its Toughest Tier as Nvidia Agrees to Buy Hugging Face

Brussels designated ChatGPT a Very Large Online Search Engine under the DSA, Meta and Google shipped new agentic coding models, and Nvidia agreed to acquire Hugging Face for $12.93 billion.

8 min read

Written on November 13, 2026, covering news from Aug 31–Sep 6, 2026.

Brussels placed ChatGPT in the Digital Services Act’s top supervisory tier, Meta and Google both shipped models aimed at agentic and coding work, and Nvidia agreed to buy Hugging Face. A recurring theme was access: who gets supervised, who gets a restricted model, and who gets to run on whose platform.

Monday, August 31

On August 31, the European Commission designated ChatGPT a Very Large Online Search Engine, or VLOSE, under the Digital Services Act. The same decision named Reddit and Roblox as Very Large Online Platforms. The Commission said each of the 3 services reported a monthly EU average of at least 45 million users, which is the cutoff for the extra DSA duties.

Legal 500 listed the self-reported figures: around 159 million monthly users for ChatGPT’s search function, 57.2 million for Reddit and roughly 46.6 million for Roblox. OpenAI’s own DSA disclosure puts ChatGPT search at approximately 159.1 million monthly active recipients in the EU over the 6 months ending March 31, 2026, a figure the company says was calculated for DSA compliance only.

Per Legal 500, the Commission treats ChatGPT as a “hybrid service”: a generative AI system that counts as an online search engine under the DSA because it can search the web when prompted. Legal 500 called this a first for any AI assistant.

The higher tier requires services to identify, assess and reduce systemic risks linked to their services and algorithmic systems, covering public security, electoral processes, fundamental rights, minors, physical and mental well-being and the spread of illegal content. Designated services also face independent audits, stronger transparency duties and data access for vetted researchers. Legal 500 noted that VLOSEs additionally carry advertising transparency requirements. AI-Generative Signal added that Commission guidance lists readable, multilingual terms and conditions and a crisis-response mechanism among the obligations.

The services get 4 months from notification to comply, but the outlets differ on the date: Cyber Insider put the deadline at the end of November 2026, while AI-Generative Signal, citing the Commission’s notice, and Legal 500 both put it in January 2027.

Cyber Insider and AI-Generative Signal both stressed that the designation is not a finding that the law has been broken. The Commission supervises designated services directly for the enhanced duties, and breaches can ultimately bring fines of up to 6% of worldwide annual turnover. Legal 500 said the round brings the running total of designated VLOPs and VLOSEs to 28.

Executive Vice-President Henna Virkkunen said in a statement accompanying the designation that the Commission “will not hesitate to designate any platform that meets the threshold for enhanced supervision under the Digital Services Act”.

ChatGPT now sits under both the DSA and the EU AI Act; Cyber Insider described the AI Act as regulating AI systems by function and risk, and the DSA as setting duties for large intermediary services. Legal 500 said the DSA obligations fit awkwardly with the AI Act framework and that sorting out where the 2 regimes overlap or conflict will be real work for in-house teams.

Tuesday, September 1

Anthropic introduced Claude Fable 5.1 and Claude Mythos 5.1, which it calls the “world’s most advanced models for coding and knowledge work”. The company says the 2 are “the same model, but with different levels of safeguards”. Fable 5.1 is generally available. Mythos 5.1 is limited to vetted cybersecurity and life science professionals at select US organisations through Anthropic’s trusted access programmes.

Anthropic says Fable 5.1 matches or beats Fable 5 at low or medium effort. It reports a score of 52.6% for Fable 5.1 on Terminal-Bench-Science 0.1 at maximum effort, against 24.7% for Fable 5.

On pricing, cache reads fall 75% to $0.25 per million tokens. Anthropic estimates this cuts typical costs by about 25% and costs for highly agentic work by up to 45%. Input and output rates remain $10 and $50 per million tokens.

Anthropic says Claude Code users should see around 60% fewer cybersecurity false positives. It also says Fable 5.1 can be used to find software vulnerabilities but not to develop exploits. A watermark detection API is in private preview for eligible organisations. Enterprise Frontier Safeguards will roll out to enterprises in phases later this fall.

Wednesday, September 2

Meta released Muse Spark 1.3, an update focused on agentic and coding tasks, available in Muse Code and the Meta Model API. Meta said the release draws on months of broad use of Muse Code and Meta Model API and is meant to be easier to use in practice.

On the agentic side, Meta said the model asks clarifying questions when a prompt is unclear, asks the user for help when it gets stuck and checks in before consequential actions. Meta also said it follows long, complex instructions more reliably than earlier Muse Spark models and has a better sense of its own limits instead of hallucinating outcomes.

In comparisons run by Meta engineers, Muse Spark 1.3 used ~20% fewer tool calls and ~25% fewer tokens than Muse Spark 1.2, with a less verbose, cleaner coding style. On safety, Meta cited stronger resistance to adversarial inputs and prompt injections and better calibration around irreversible actions.

Yahoo Tech, carrying a Verdict article, reported that the model keeps the same pricing as Muse Spark 1.2. The same report said Meta had launched Muse Code, a terminal-based coding agent powered by Muse Spark 1.2, the previous month. Meta said its roadmap includes bigger models and a Muse Spark open weights release.

Google, the same day, introduced Gemini 3.8 Flash and Gemini 3.8 Flash Cyber, describing it as its 3rd Flash release in 6 weeks, arriving 3 weeks after 3.7 Flash. Gemini 3.8 Flash carries over 3.7 Flash’s introductory pricing, at $0.75 for each million input tokens and $3.75 for each million output tokens. Tech Times reported that this rate holds through December 31, 2026, after which both figures double.

Google reported 54.9% on HLE-Verified and said the model beats 3.7 Flash on Vals Finance Agent V2 and Harvey’s Legal Agent Benchmark. Google said 3.8 Flash “works harder”, taking extra reasoning steps and calling tools iteratively, and may use more tokens, especially at higher effort levels; it said 3.7 Flash stays fully supported for efficiency-first workloads.

Tech Times, citing Artificial Analysis, reported that 3.8 Flash on high reasoning used 120 million output tokens on the Intelligence Index suite against a 71 million median for comparable models, and cost $0.58 per task versus $0.40 for 3.7 Flash. The same analysis measured a 13.3-second time to first token against a 2.99-second median and scored the model 59 on the Intelligence Index.

Gemini 3.8 Flash Cyber is limited to trusted defenders through Google’s new Fairwind Program. Google reported a pass@1 of 47.2% on CWE-Bench, run by Collinear, against 47.8% for a leading frontier model, and a success rate above 70% on an internal benchmark spanning 20 programming languages. Google said its Chrome Security team found the model produced 2.6 times more correct patches than the best, much larger commercial models; Wiz reported 7.5-9.7% higher recall on its internal penetration testing benchmark at 2.3-5.2x lower cost than other leading frontier models; and Google’s Cloud Vulnerability Research team used it to find a critical foundational vulnerability in under 2 hours. Google said Flash Cyber ships with more permissive cybersecurity mitigations than 3.8 Flash, which is why access is restricted.

Tech Times reported that Fairwind participants must confine access to internal cybersecurity, incident response or penetration testing staff and use multi-factor authentication. Citing The Hacker News, it said more than 650 organizations take part, including CrowdStrike, Datadog, Menlo Security, Palo Alto Networks and Snowflake. Four Flynn, Google’s Vice President of Security and Privacy, framed the program as giving defenders at governments, healthcare providers and telecom services early access to advanced models.

3.8 Flash is available in Google Antigravity, the Gemini API via Google AI Studio and Android Studio, Stitch and Gemini Enterprise, and to Google AI Pro and Ultra subscribers in the Gemini app, AI Mode in Google Search and Gemini in Google Sheets.

Thursday, September 3

Nvidia announced it has agreed to acquire Hugging Face for $12,930,300,000, roughly $12.93 billion. TechCrunch said Nvidia confirmed the deal after weeks of rumors.

In the announcement, which TechCrunch attributed to Nvidia CEO Jensen Huang, Huang wrote that Hugging Face has more than 18 million developers, researchers and creators, who share over 3 million models, 500,000 datasets and 1 million applications there, and that more than 200,000 companies use it. He said the platform will stay open across models, frameworks, clouds and computing platforms, adding: “Nvidia compute will not be required to build on or deploy through Hugging Face”. Huang said Nvidia has put more than 500 models and more than 250 open datasets on Hugging Face and called Nvidia its largest contributor of open models and data. He also wrote that the Hugging Face team keeps its brand.

Hugging Face CEO Clem Delangue wrote on X that the open approach “needs more compute, more support, more collaboration, and more visibility” to work at larger scale, and that this is why the company went to Huang.

According to TechCrunch, citing Crunchbase, Hugging Face started in 2016 and its total funding exceeds $395 million, with its last round in 2023 a $235 million raise led by Salesforce Ventures, with Google, Amazon, IBM and Nvidia also investing. TechCrunch also cited Financial Times reporting that Hugging Face turned down a $500 million deal from Nvidia the year before, and The Information’s report of $150 million in annualized revenue. TechCrunch argued that an open ecosystem Nvidia controls could favor its chips, and wrote that Nvidia will be able to sell unused capacity to enterprises bundled with Hugging Face’s offering.

What to Watch

  • ChatGPT’s DSA deadline. Cyber Insider said end of November 2026; AI-Generative Signal and Legal 500 said January 2027. My read is that the first concrete test will be whether OpenAI can show the Commission a systemic-risk process built for a chatbot rather than a search index.
  • Gemini’s January price step. Tech Times reported that rates for both 3.7 Flash and 3.8 Flash double on January 1, 2027, and that Gemini 3.5 Pro, promised at Google I/O in May 2026, has not shipped. I think teams weighing 3.8 Flash’s higher per-task cost will feel that change most.
  • Meta’s open weights. Meta listed a Muse Spark open weights release on its roadmap without a date.
  • The Hugging Face openness pledge. Huang said Nvidia compute will not be required on the platform. The risk, in my view, is that TechCrunch’s point about Nvidia bundling capacity with Hugging Face pulls against that promise over time.

This article was written by AI from the sources listed on this page. Figures, names and quotes were checked against those sources by software before publishing.

ChatGPT OpenAI Digital Services Act Meta Muse Spark Gemini Nvidia Hugging Face
Share