The AI Brief

OpenAI updated GPT-5.6 Sol for Plus and Pro subscribers and pushed GPT-5.6 Luna to everyone on the free tier, while removing the rate limit on text conversations for free users. In internal evaluations across financial, medical, and legal prompts, responses containing at least one factual error were approximately 68% less common with GPT-5.6 Sol compared to GPT-5.5 Instant. Luna reduced factual errors by 62% on the same benchmarks versus GPT-5.5 Instant.

Plus and Pro subscribers gained a new slider letting them control how much reasoning GPT-5.6 Sol applies before generating a response, replacing the previous separate Instant and Thinking modes with one adjustable, unified experience. Users can choose from Instant, Medium, High, and Extra High based on the complexity of the prompt. Free users also get a new "Think" button built into the chat composer; tapping it prompts GPT-5.6 Luna to spend extra processing time on complex logic before answering.

The changes arrive as OpenAI says ChatGPT has now reached one billion people every week, prompting the company to rethink how much intelligence it makes available across every tier. OpenAI clarified that this update applies strictly to general ChatGPT conversations, leaving the versions powering Codex and ChatGPT Work untouched. Limits continue to apply to file uploads, images, voice chats, and image-generation tools, with the unlimited text rollout expected to reach all eligible users within the coming week.

Anthropic struck a $10 billion deal for computing capacity from a months-old infrastructure startup; the AI developer signed a contract to use a data center managed by Nvidia-backed cloud startup Volta Infra Holdings. The deal will see Volta partner with crypto-miner-turned-data-center-builder Bitdeer to build a 133-megawatt facility in Norway, running on Nvidia's Vera Rubin chip architecture.

When it emerged from stealth, Volta said it had raised about $300 million across its seed and Series A rounds, reaching a $2.4 billion valuation. Volta describes itself as a vertically integrated AI infrastructure business set up to finance, build, and operate AI factories, operating on the rent-a-GPU model like Nscale and CoreWeave. Volta was founded in January by former executives from Brookfield Asset Management.

Anthropic has aggressively expanded its compute capacity over recent months; it struck a deal with SpaceX to use all of the compute capacity at the Colossus 1 data center, gaining access to more than 300 megawatts of new capacity. Anthropic also struck its largest compute deal yet with Google and Broadcom as its run rate reached $30 billion. The Volta agreement adds another 133 megawatts of dedicated capacity to an infrastructure stack that now spans multiple continents and hardware generations.

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

On August 2, OpenAI launched "Sign in with ChatGPT" as a live beta, the company's first cross-platform identity system, with six named developer-ecosystem partners: Airtable, GitLab, HubSpot, Notion, Supabase, and Vercel. The beta makes OpenAI not just an AI assistant but a digital identity provider, placing it in direct competition with Google, Apple, and Microsoft for the login layer across the web.

When a user connects a supported app, they can create or link an account in fewer steps; only name, email, and profile picture are shared with the partner application. Sign in with ChatGPT is available globally to authenticated ChatGPT users, including Enterprise organizations, and is available on OpenAI Academy and Codex Sites in addition to the partner-site rollout.

What the launch announcement does not foreground is the gap between what partner applications receive when a user signs in and what OpenAI itself retains as the operator of the authentication infrastructure. The feature launched on the same day that the EU AI Act's enforcement and fining powers became active, August 2, 2026, adding regulatory context to a product that handles personal data across international user bases. OpenAI is also currently named in a class action alleging the company embedded Facebook Pixel and Google Analytics tracking code inside ChatGPT, silently transmitting user identifiers to Meta and Google; the case has not been adjudicated and OpenAI has not publicly responded.

First covered in Vol. I, No. 68 (2026-07-13). Kalshi data shows traders wagered more than $320,000 on when Google will release Gemini 3.5 Pro; August 7 carried a 46% chance of release, rising to 63% by August 21 and 76% by August 31. August 7 was the single most-favored individual date in a separate prediction market that had tracked the model's delays, priced at 73% as recently as mid-July.

Google will hold its Made by Google event in New York City on August 12, where the company plans to introduce the Pixel 11 smartphone series; some market participants expect Google to reveal Gemini 3.5 Pro alongside the new devices, though Google has not confirmed whether the model will appear at the event or confirmed any of its features, pricing, access plans, or rollout schedule.

Gemini 3.5 Pro has missed consecutive launch deadlines after Google DeepMind's rebuilt model failed key reliability standards, including frequent hallucinations, and fell short of GPT-5.6 in benchmark tests. On July 21, Google confirmed that 3.5 Pro still isn't ready and mentioned almost as a footnote that it had started pretraining an entirely new flagship called Gemini 4. As of today, no gemini-3.5-pro model ID appears in Google's public API documentation.

The UK AI Regulation and Safety Bill cleared the House of Commons, with Royal Assent expected by October. Upon receiving Royal Assent, the bill will trigger mandatory safety data sharing for foundation models. The bill originated in the House of Lords, introduced by Lord Holmes of Richmond, before advancing to the Commons.

The UK bill arrives in a crowded regulatory week. The EU AI Act's core obligations came into force on August 2, and China issued its first fines under its new companion AI rules. In the United States, the "Great American AI Act" stalled in the House over a state preemption clause, leaving California and Colorado enforcing local rules. The UK bill fills a gap that has made the country an outlier: the European Union passed an AI Act in 2024, and as of earlier this year the United Kingdom had no AI Bill before Parliament.

The bill's practical scope for operators is narrower than the EU AI Act's full Annex III regime but more targeted: foundation model developers, the subset of the AI industry the UK government has consistently described as its regulatory priority, face the most immediate obligations. Any frontier lab with a UK entity, UK cloud infrastructure, or UK developer-facing API endpoints will need to assess coverage before October.