Today's Headlines

  • OpenAI says ChatGPT Ads has reached a $1 billion annualized revenue run rate — self-service buying opens across India, Europe, the Middle East and North Africa
  • The European Commission designates ChatGPT a Very Large Online Search Engine under the DSA — Reddit and Roblox are designated as platforms
  • DeepSeek publishes the first multimodal experimental model in its V4 family as open weights under the MIT licence
  • Nvidia takes a $3.5 billion stake in MediaTek, pulling custom AI chips onto its own fabric through NVLink
  • Claude Code's weekly limit rises 25 percent permanently on September 14 — and still sits about 17 percent below today's cap

Today's three stories all turn on ChatGPT. In a single day the same product took three positions at once: the one selling advertising, the one designated a search engine and brought under regulation, and the one being chased by open weights.

The first is about how the money comes in. OpenAI treats advertising as a pillar that supports the free tier, and it opened the buying window to new regions.

The second is about which box the product sits in. The European Commission looked at the behaviour of searching the web in order to answer, and it sorted ChatGPT onto the search engine side.

The third is about who is doing the chasing. The weights DeepSeek released carry, on their own model card, scores that beat a closed frontier model on part of the agentic work that involves vision.

Selling, being regulated and being chased usually land on three different companies. What marks this Tuesday is that all three landed on one product within a single day.

Today's Top Three

1. OpenAI says ChatGPT Ads has reached a $1 billion annualized revenue run rate, and opens self-service buying across India, Europe, the Middle East and North Africa

OpenAI said on August 31 that ChatGPT Ads, its advertising business inside ChatGPT, had reached $1 billion in annualized revenue run rate less than 200 days after launch.

Start with what the figure measures. An annualized revenue run rate extends the current pace of sales across a full year. It is a projection of today's pace rather than a year of sales already banked, and the distinction matters when reading the number.

The same announcement moved the sales channel. OpenAI wrote that self-service buying through Ads Manager would open later that same day across India, Europe, the Middle East and North Africa.

Until now, buying ran mainly through OpenAI's own sales team and through agency and technology partners, with Ads Manager opening a door for smaller businesses when it arrived in May. Ads run in more than 40 countries, technology and measurement partners number more than 50, CPC and outcome-optimized bidding account for the majority of campaigns, and Pixel and the Conversions API carry the measurement.

The results OpenAI cites are individual cases. A 3x return on ad spend over 28 days comes from one ecommerce advertiser, and the finding that more than 80 percent of ad-driven traffic was new customers comes from one technology partner. Read either as a platform-wide average and the number stretches well past what was reported.

The substantive point is that the free tier now rests on advertising. OpenAI places ads alongside consumer subscriptions, enterprise offerings and usage-based APIs as one of four pillars, and describes the ad-supported free tier as what keeps ChatGPT open to more than a billion weekly users.

It also states that ads are always clearly labelled and kept separate from answers, that advertising leaves the answers themselves untouched, and that advertisers get no access to private conversations. Each organisation will want to test that against its own yardstick; at this stage the assurance rests on the provider's own statement.

Our August 28 edition covered the start of ad delivery to the free and Go tiers of ChatGPT in India. Today's announcement continues that line: India has moved from the start of delivery to the opening of self-service buying.

2. The European Commission designates ChatGPT a Very Large Online Search Engine under the DSA, with Reddit and Roblox designated as platforms

The European Commission designated ChatGPT a Very Large Online Search Engine (VLOSE) under the Digital Services Act on August 31, and designated Reddit and Roblox Very Large Online Platforms (VLOP).

Designation starts with self-declaration. All three services reported crossing the threshold of 45 million average monthly users inside the EU.

The reasoning is the heart of this one. The Commission describes ChatGPT as a hybrid service that responds to user prompts and queries and searches the web in the process, and concludes that it qualifies as an online search engine under the DSA.

Reddit and Roblox are online platforms because they disseminate third-party content to the public. Two services can look like the same generative AI chat, and whether a web search sits inside the loop decides which box they land in. That distinction bears directly on how a company positions its own service.

The clock runs four months, to January 2027. Within that window the services must meet added obligations to assess and mitigate systemic risks covering the spread of illegal content, harm to minors, users' physical and mental wellbeing, fundamental rights, electoral processes and public security.

What was settled today is the designation itself, with the compliance deadline set for January 2027. The designation also gives the Commission investigatory powers reaching into how the services function.

Supervision is shared between the Commission and national Digital Services Coordinators. Ireland's Coimisiún na Meán handles ChatGPT and Reddit, and the Netherlands Authority for Consumers and Markets (ACM) handles Roblox. This brings the total number of DSA designations to 28.

3. DeepSeek publishes the first multimodal experimental model in its V4 family as open weights under the MIT licence

DeepSeek published the weights of DeepSeek-V4-Flash-Vision-Exp, the first multimodal (vision-language) experimental model in its V4 family, on Hugging Face on August 31 as open weights under the MIT licence.

The model carries roughly 305 billion total parameters, distributed in fp8 and int8. The Hugging Face entry was created at 06:16 UTC that day and last updated at 12:23. The Exp in the name stands for experimental, and the release is presented as such.

What tells you most about this release is that the model card itself lines the model up against Opus-4.8. It comes out ahead on two of the four comparisons: Agents Last Exam, a multimodal agentic benchmark, at 27.3 against 25.7, and ZeroBench Pass@5 at 35.0 against 34.0.

It comes out behind on the other two. Terminal Bench 2.1 reads 83.9 against 85.0, and ApexBench Pass@1 reads 36.5 against 39.4. Two ahead and two behind is the shape of it, and reading the release as drawing level across the board claims more ground than the numbers cover.

On text-only agentic work, DeepSeek says the model holds parity with the previous DeepSeek-V4-Flash-0731 (DeepSWE at 59.3 against 54.4, Toolathlon-Verified at 75.9 against 70.3). Every figure above comes from the comparison table on the model card.

The consequence worth noting is that MIT-licensed weights are now an option for agentic work that involves vision. For organisations that want to run screen-reading and document-reading automation in house rather than through an outside API, the ground available for evaluation just widened.

Elsewhere

Policy and regulation

  • Internal chat logs are being cited as evidence in the copyright suit brought against Anthropic by Sony Music and others. This is the same case that led our August 30 edition; what is new today is the complaint's citation of internal chats about mirroring a pirate library, including a reply reading "zlibrary my beloved", and the addition of co-founder Benjamin Mann and chief executive Dario Amodei as individual defendants. Both remain allegations in the complaint at this stage. — "Zlibrary my beloved": Anthropic staff chats extolling piracy cited in Sony suit (Ars Technica)
  • Debian voted to adopt a policy that allows AI-generated code, treating disclosure as a recommendation rather than a requirement. As an example of an open-source project deciding to keep one standard for all contributions, it is a useful reference when drafting an internal code-contribution policy. — Debian won't ban AI code from its Linux distribution (The Verge)
  • Instagram is limiting the reach of undisclosed AI-generated profiles, and has renamed the label from "AI creator" to "AI-generated profile". Labelling carries no penalty while going unlabelled costs distribution, and ordinary uses of AI such as photo editing and caption writing fall outside the rule. — Instagram puts new limits on undisclosed AI profiles (TechCrunch)
  • New York Governor Kathy Hochul set out the thinking behind her one-year freeze on data centres, explaining that five pending applications in St. Lawrence County alone add up to the power demand of a single nuclear reactor. She also said the state uses AI to review its own regulations, saving more than 1.5 million hours across the first 50. — New York Governor Kathy Hochul thinks AI should be 'less evil' (The Verge)
  • An appellate brief dated August 31 was entered on the docket in the suit over the removal of the US Register of Copyrights. The brief recounts that the removal followed a day after the Copyright Office published Part 3 of its report on the use of copyrighted work in AI training. We were unable to obtain the docket itself; our basis here is the excerpt of the brief returned by a search feed. — Shira Perlmutter v. Todd Blanche (CourtListener)

Business

Products and features

Research

  • A study at Lille University Hospital in France and elsewhere found that 44 percent of medical residents accepted a disease name that AI had invented. Eighteen of 41 residents folded the fictitious condition "neurocadmium syndrome" into their differential diagnosis. Among those with six months or less of neuroradiology experience, 69 percent were taken in, while everyone with more than six months caught it. The same AI assistance lifted diagnostic accuracy from 52 to 61 percent, so benefit and fragility arrive together from one tool. — Residents accepted an AI-invented disease name in 44% of cases (ITmedia NEWS)
  • A survey reports that 87.7 percent of small and mid-sized Japanese firms feel strong unease about entering confidential information into generative AI, while 11.4 percent have technical controls in place. The survey is Yamanex Corporation's own, released through PR TIMES with n=114, and the company begins selling its "Buchi AI" internal AI environment on September 2. A separate survey found that 43.3 percent of 150 IT decision-makers across five prefectures in Japan's Chugoku region had neither the people nor the structures, in-house or outside, to drive AI and digital transformation. Both samples are small vendor surveys, so the direction reads more safely than the levels. — 87.7% of SMEs using generative AI report strong unease about entering confidential data (ASCII.jp)

Source: selected by the editorial desk from the AI news inbox (collected September 1, 2026 — 25 items, 4 primary and 21 secondary).