The AI Brief

OpenAI banned a cluster of Russian ChatGPT accounts on August 25 after an investigation traced AI-generated social media content back to a coordinated campaign built around the International Burke Institute (IBI), a self-described "expert community" claiming a street address in Israel. OpenAI recently banned a cluster of ChatGPT accounts originating in Russia that were being used to promote the IBI, a self-described "expert community" based in Israel. The operators accessed ChatGPT via VPNs to circumvent OpenAI's Russia restrictions, routing their connections through VPNs and using ChatGPT to draft social media content that appeared on Substack, Telegram, X, Facebook, and LinkedIn.

In a review of a sample of 36 articles linked to experts on the IBI website and published between September 2025 and May 2026, 34 of the articles were copied from elsewhere on the internet. In one particularly glaring example, a supposed migration policy expert named in a misattributed paper was actually a food sciences professor from Australia. The site also ran a sovereignty index that relied on an internal IBI calculation that was uniformly critical of Western countries at odds with Russia over its invasion of Ukraine. OpenAI rated the campaign Category Three on the Brookings Breakout Scale.

While the immediate impact of the operation appeared limited, OpenAI said the "significance of the operation lies less in the audience it reached, however, than in the infrastructure it had built." Posts produced with ChatGPT steered readers toward what looked like a legitimate research body, a tactic OpenAI said demonstrates how AI tools can help bad actors create an appearance of institutional credibility, bury the true source of preferred narratives, and lay groundwork for operations that could grow considerably larger.

Lambda Inc., an AI cloud-computing provider backed by Nvidia, is in talks to raise as much as $3 billion in a round that could tee it up to go public next year, according to people familiar with the efforts. The company has discussed raising the money at a valuation of as much as $12 billion or more, the people said, asking not to be identified because the discussions aren't public.

Lambda is expected to generate more than $1.5 billion in revenue in 2026, and the round would come roughly nine months after its $1.5 billion Series E, as rival neoclouds CoreWeave, Crusoe, and Nscale all chase similarly large raises or listings. On August 12, Lambda priced a $926 million senior secured term loan B facility, which it described as the first investment-grade-rated term loan B financing executed by a private neocloud. The new equity round, if it closes, would layer on top of that debt rather than replace it.

Nvidia is an investor, a supplier, and indirectly a source of demand, an arrangement it has replicated across the sector. The wider question for any neocloud is what happens when the capacity crunch eases, as these businesses earn their margins on scarcity, and the same customers renting GPUs today are building their own silicon or negotiating directly with hyperscalers. While Lambda would not be the first neocloud to launch an IPO, CoreWeave launched in 2025, it comes into a market with several analysts anticipating that AI demand will cool, and its IPO could serve as a canary in the coal mine.

OpenAI officially retired the o3 model family from ChatGPT on August 26, 2026, ending a 90-day sunset period that began with the May 28 announcement. The model benchmarked at 96.7% on AIME 2024, 87.7% on GPQA Diamond, and a 2,727 Codeforces Elo and will quietly disappear from the model picker for paid subscribers. One notable carve-out: o3-pro, the higher-compute variant, will remain available in ChatGPT for Pro, Team, Enterprise, and Edu subscribers after today.

OpenAI has retired GPT-4o, GPT-4.1, o4-mini, GPT-4.5, and now o3 from ChatGPT, more models in 2026 than in all prior years combined, with the model lifecycle compressed from roughly 18 months to closer to six. The API shutdown of o3 on December 11, 2026, replaced by GPT-5.6 Sol, will be the true stress test for enterprise migration readiness. Note that today also marks the hard shutdown of the Assistants API, covered in Vol. I, No. 92.

OpenAI reinstated the five-hour rolling usage cap on Codex and ChatGPT Work for Plus subscribers on August 25, framing it as compute load-smoothing. Engineering lead Thibault "Tibo" Sottiaux said the cap lets OpenAI "smoothen the load on our compute," and confirmed Pro $100 and Pro $200 subscriptions will not have the five-hour limit re-enabled "for the upcoming months." The cap's return was cited by some builders as reason to re-evaluate Codex against Claude Code. First covered as an announcement in Vol. I, No. 92.

Gemini 3.5 Pro is still unreleased as of August 23, 2026, more than three months after Google announced it at I/O on May 19. It has now missed three targets, June, mid-July, early August, and has no model ID, no pricing, and no launch date. The most recent entry in the Gemini API release notes is August 13, 2026, when Gemini 3.7 Flash went generally available. The official Google DeepMind models page lists Gemini 3.1 Pro as the current Pro model and carries a "3.5 Pro coming soon" marker.

The story behind the delay traces to engineering decisions made months ago. Reports cite persistent coding and reliability problems, senior researcher departures, and a possible complete retraining from the foundational pre-training phase due to a "structural problem." In a single week in June, Google lost Gemini co-lead Noam Shazeer to OpenAI and Nobel laureate and AlphaFold coinventor John Jumper to Anthropic. Google, which acquired DeepMind in 2014, has not unveiled a frontier model since early 2026.

First covered as an ongoing delay in Vol. I, No. 90, with daily updates since. The delay has now extended past the Made by Google hardware event (August 12), the period covered in No. 91 at day 97, and every other window prediction markets assigned meaningful probability. Google has not commented publicly on a revised launch date.

Anthropic said on August 21, 2026 that Claude Mythos 5, the cyber-capable model it has limited to vetted defenders since April 2026, is now running vulnerability scans in Claude Security for Enterprise customers. The same announcement lays out plans to bring Mythos 5 into security partners' products, commits $35 million in credits to a new fund for open-source defense, and previews an expansion of the company's Cyber Verification Program toward Mythos-class access.

The scanning process connects to a GitHub repository, traces data flows across files, and provides users with findings categorized by CWE, along with confidence and severity ratings and suggested patches. Anthropic explained: "Claude Security uses Mythos 5 to scan code you own, and returns detailed findings rather than raw outputs without exposing the model itself," meaning defenders can access Mythos 5's capabilities without the model becoming accessible to those who might misuse it. Anthropic also launched the Defender Advantage Fund (0xDAF), putting $35 million in Claude credits toward organizations that help open-source maintainers secure their projects.

Grants will focus on patching live vulnerabilities in widely used projects, automating scanning and patching in ways other projects can replicate, and helping projects pursue security approaches that make them resistant to whole classes of attack. Anthropic is starting with a small number of larger pilot grants to learn what works and scales best, and will share details on initial recipients in the coming weeks. Traffic on Mythos-class models carries a 30-day retention requirement, which Anthropic says is used for safety purposes and not training.

Disclosure: Claude, which generates this brief, is built by Anthropic.