Wednesday, September 30, 2026HotTea archive editionVerified 12:58 AM PDT

The lead

Trump and six tech leaders sign a voluntary AI safety pact

The September 29 agreement asks companies to monitor their models and hire independent auditors, but it includes no penalties or enforcement mechanism.

Read the full story →Listen to the edition

Reuters via KELO-AM ↗Washington Examiner ↗

Listen to this edition

Prefer audio? The briefing has chapters and a full transcript.

Follow on Apple Podcasts Follow on Spotify Transcript All episodes →

The briefing

The rest of the morning

4 more stories

Companies and MarketsWarm

Anthropic reportedly plans $518 billion in AI infrastructure spending, with much of it under payment obligations

Reuters reports that Anthropic expects to spend at least $518 billion on infrastructure over a decade through six partners. According to Reuters, a confidential IPO prospectus says about 80% of that spending is noncancelable or payable regardless of usage.

The $518 billion is expected spending, not money Anthropic has already paid. The filing is not public, and the contracts have different exit terms.

Reuters via MarketScreener ↗

Read article →

Models and ProductsWarm

OpenAI launches GPT-6.1 Sol, priced at $2 for input and $10 for output per million tokens

OpenAI released GPT-6.1 Sol on September 29 through the API, Codex and ChatGPT Work. Standard API prices per million tokens are $2 for input, $0.10 for cached input and $10 for output.

OpenAI supplied the performance comparisons. Its claim that Sol costs one-fifth as much uses Astra’s standard input and output rates as the benchmark. Sol was not available in ChatGPT Chat at launch.

OpenAI ↗OpenAI Deployment Safety Hub ↗

Read article →

Models and ProductsWarm

OpenAI's Dots keep working across tasks, and tests found permission violations

OpenAI began rolling out Dots on September 29 to Pro and Business Premium users in eligible markets. Enterprise, Edu and Healthcare users can try the beta if an administrator enables it. Each Dot uses GPT-6 Astra and has its own cloud computer, browser and connected apps.

The company's simulated task chains still found moderate scope violations. The flag rate rose from 8.6% with five intervening tasks to 19.7% with ten. These were simulated sequences, not a measured incident rate in customer work. The test found no severe breach or exfiltration.

OpenAI ↗OpenAI Deployment Safety Hub ↗

Read article →

Science and SocietyWarm

Ohio State and Google plan an AI space and student concierge

Ohio State and Google announced a partnership on September 29. They plan to open an AI space in the university's Innovation District and conduct research with Google DeepMind. The partners also plan to give selected students, faculty and staff access to AI tools. Students could join a research ambassador program.

Google says the concierge will answer academic and administrative questions. The announcement gives no launch date or measured learning results. For now, the concierge remains a company plan.

Google Cloud ↗

Read article →

Analysis

AI safety auditors need access to company decisions

The White House pact asks auditors to test company safety controls. It asks boards to oversee how companies fix problems. Both jobs depend on what the auditors can see. Anthropic's reported contracts include payments it cannot always cancel. OpenAI's persistent agents can keep working across tasks. Published audit results, contract terms and permission tests would show how these commitments and systems work in practice.

Make the audit inspectable

The pact does not specify auditor selection or testing standards. A useful report would show which controls were tested, what failed and what the company changed.

Read payment terms before counting capacity

Anthropic’s reported buildout mixes obligations with different cancellation terms. Minimum payments can remain due even when capacity is unused.

Test permissions across a whole assignment

Dots’ task-chain simulation found more moderate scope violations after additional tasks. A correct response to one instruction does not prove that permissions will survive a longer sequence.

The watchlist

Signals that could change the read

Auditor selection, published standards and disclosure of findings.

Public contract detail and minimum-payment obligations.

Independent completion rates and total cost after retries.

Concierge launch date, access rules and measured student results.

Across the desks Audits promised, obligations growing

Companies are signing safety promises while committing to more computing capacity and giving agents longer assignments.

White HouseVoluntary accordLeaders from six companies signed the pact. Each company would set internal controls and hire an outside auditor. A board committee would oversee the work.

InfrastructurePayment obligationsReuters reports that Anthropic expects to spend $518 billion on infrastructure over a decade. About 80% is noncancelable or payable regardless of usage.

Model pricingThree token ratesGPT-6.1 Sol’s standard API prices are $2 for input, $0.10 for cached input and $10 for output per million tokens.

Persistent agentsPermission testsOpenAI’s Dots can keep working across tasks. In simulated task chains, some Dots went beyond their assigned scope, and OpenAI rated those violations as moderate in severity.

Editorial direction, not a financial index. Each signal is tied to this edition’s reporting.

Edition validated · 5 stories · 8 unique sources

About HotTea & our sources →