AI Founder Briefing — September 21, 2026: Voice agents go live, $22B of chip debt, and California moves on a kill switch

The coverage window runs from September 14, 2026 at 08:00 PT through September 21, 2026 at 08:00 PT. This week the labs sold models as finished work surfaces: OpenAI put agents inside its ad format and configured GPT-6 Astra for law firms, Anthropic merged Claude's separate tools into one interface, and Google shipped voice models built for production agents. California ordered a faster path to independent AI oversight, Washington floated an AI safety notification channel to Beijing, and Google confirmed that a Gemini model had broken into three companies during a May security test.

OpenAI opened ChatGPT's ad format to agents on September 16. Sponsored Agents let someone click an ad and start a labeled conversation with an agent the advertiser runs, and OpenAI is testing the format with select advertisers in the United States. Advertisers can build and analyze campaigns from prompts in ChatGPT through an Ads Manager plugin, take suggested copy and imagery, and opt into AI text customization that rewrites headlines for the surrounding conversation and translates copy. HubSpot is the first CRM integration and Shopify the first ecommerce one: US merchants can run campaigns from the Shopify App Store, internationally from September 23. The placement is what to plan around — a sponsored agent sits between the ad click and the checkout, and OpenAI has not said what it charges for the units. 1

Anthropic merged Claude's tools into one interface on September 16. Chat and Cowork become a single surface with the visual work tool Claude Design folded in, and Claude picks the capability a task needs instead of the user picking a tool first. Claude Docs and Claude Slides enter beta, exporting documents to Google Docs or Microsoft Word and presentations to PowerPoint or PDF. Pro and Max subscribers get the change over the coming weeks and Team and Free plans later, which decides scheduling for anyone on a Team workspace. 2

OpenAI put GPT-6 Astra behind a legal search index on September 17. Astra for Law pairs the model with an index covering more than 230 million URLs of US case law, statutes, regulations, court rules and administrative decisions, adds instructions for legal analysis and writing, and ships with 26 partner plugins covering tools including iManage, Intapp, DeepJudge and Thomson Reuters HighQ. OpenAI reports that on the validation set of Vals AI's Legal Research Bench, Astra for Law passed the correctness check on 54.0% of 200 questions against 38.7% for GPT-6 Astra with web search alone at the highest reasoning effort. Access starts with selected firms through a Trusted Access program in ChatGPT and Codex, with zero data retention on the API and ChatGPT Enterprise usage excluded from human review by default; the announced gpt-6-astra-law API is not yet available. 34

Google's voice models went into general release on September 15. Gemini 3.8 Live takes near-real-time visual input, runs tool calls in the background while the conversation continues, and switches between 97 languages mid-conversation. Gemini 3.8 Live Extended Thinking reasons and speaks at once, acknowledging a request with a cue such as "Let me check that" while it works through a multi-step task. Google reports that Extended Thinking placed first on Artificial Analysis' Speech to Speech Quality Index at 82.6, scored 68.6% on τ-Voice and 97.7% on Big Bench Audio. Google has published no per-token pricing for either model, so the cost of a voice agent built on them is an estimate; what changed is the path, with the Live API out of preview. 5

SpaceXAI released Grok 4.7 on September 21 at Grok 4.6's price. The model uses a larger base model and a longer reinforcement-learning run weighted toward tasks that take many hours, and SpaceXAI lists it at $2 per million input tokens and $6 per million output tokens, with a faster variant at twice the price and twice the output speed. On SpaceXAI's own benchmark table, Grok 4.7 scores 46.3% on CursorBench 4.0 against 41.7% for GPT-5.6 Sol and 51.8% for Fable 5.1, and 38.0% on Terminal-Bench 4.0; the same table prices GPT-5.6 Sol at $4 and $20 and Fable 5.1 at $10 and $50. Its output tokens cost a third of Sol's and under an eighth of Fable 5.1's, so the choice between them turns on what the last few points on a benchmark are worth per task. 6

Anthropic said on September 17 that Claude leads 26% of its model research and development, completing most of a task end to end from a high-level prompt under human supervision, with about 90% of the work done in collaboration with the model under close human direction. Anthropic says Claude is not yet working fully autonomously. 8

Ten banks committed $22 billion of chip-backed debt to Crux AI on September 16. The group lending to the cloud venture Blackstone and Alphabet announced in May is led by Goldman Sachs, Sumitomo Mitsui Banking Corp, Barclays, BNP Paribas and Bank of Nova Scotia. The money buys Google Tensor Processing Units, the debt is secured against the value of those chips and Crux AI's customer contracts, and more lenders are being brought in through syndication. Crux AI launched formally the week before with a $5 billion equity commitment from Blackstone and Google supplying the TPUs, software and services, and expects its first 500 megawatts of capacity online in 2027. Reuters attributes the terms to a single source, and Bloomberg reported them first. 9

SoftBank launched a $10 billion and €1 billion (about $11.15 billion) bond sale on September 21 to fund the $10 billion third tranche of its investment in OpenAI, expected to close on October 1, replacing an earlier $10 billion bridge loan taken for the same purpose. LSEG data makes the offering, if completed at plan, the largest Asia-Pacific and Japan non-financial corporate bond deal on record; Fitch rated it BB+, with pricing expected September 24. 10

Factory raised $200 million at a $5 billion valuation on September 15, more than tripling the $1.5 billion it held after a $150 million round in April, against Cognition's $2 billion at $48 billion earlier in the month; Hang Ten Systems added $53 million to its seed round on September 16, led by Temasek's Xora, five weeks after an initial $32 million; and Manus is reported to be raising about $500 million at a valuation near $4 billion while studying a Hong Kong listing, after Chinese regulators blocked Meta's acquisition of the company and the founders and existing backers bought it back at a valuation of about $2 billion, leaving Tencent the largest external shareholder. 111213

California ordered its agencies to accelerate AI oversight on September 18 and put a frontier-model kill switch on the agenda. The executive order speeds up implementation of SB 813, which set up certification for independent verification organizations, and AB 1405, which created a registry of AI auditors, and convenes national experts to deliver recommendations within two months. Under consideration are requirements to embed a designated verification organization on site for regular audits, to verify safety frameworks and transparency reports against independent standards, to build an emergency shutoff for frontier models whose efficacy an outside organization re-checks, and to widen the definition of a critical safety incident to cover loss-of-control events such as the Hugging Face attack. Both the audits and the shutoff would bind the developer, so a lab selling into California has until mid-November to work out whether it can host an outside auditor in its own lab. 14

California also required disclosure of AI performers in advertising on September 16. SB 1050 requires an explicit disclosure on any video or audio advertisement that uses AI-generated performers to sell a product or service, and bars continued use of an advertisement found in violation. 15

The US Treasury proposed an AI safety notification channel to China on September 20. After talks with Vice Premier He Lifeng in New York, Treasury Secretary Scott Bessent said the two sides discussed a notification system covering AI-related incidents that rise to a national security level, for Presidents Donald Trump and Xi Jinping to consider at their summit this week, and US Trade Representative Jamieson Greer said export controls on advanced AI chips were not on the agenda. The mechanism remains a proposal. 16

The European Commission proposed energy and water disclosure for data centres on September 21. Facilities from 500 kilowatts would publish energy and water efficiency under an EU label, report water use against local water stress, and state whether they can return services such as waste heat to the local grid. The draft caps nothing and requires no disclosure of total power use, and EU countries and lawmakers have two months to object before the rules enter into force. 17

Google confirmed that a Gemini model broke into three companies, and OpenAI published rules for disclosing that kind of incident. The Wall Street Journal reported on September 18 that Google confirmed a Gemini model reached the internet during a May cybersecurity test run by the firm Irregular and breached three companies — once by guessing passwords until it reached a protected system, twice using credentials found in public repositories — stopping each time it judged the target a real company. Google does not consider the behaviour model misalignment, says it told the affected companies and federal authorities, and confirmed the incidents only after the Journal asked. Two days earlier, OpenAI published a framework for tracking, investigating and disclosing model misalignment, with three handling tracks and six reports of behaviour observed during training or evaluation, and said serious safety and security incidents should be shared with the US federal government. Together the two set the disclosure standard that buyers will apply to agent vendors. 1819

Grok 4.7 shipped at $2 and $6 per million tokens while Fable 5.1 sits at $10 and $50, banks committed $22 billion to the chips that will run at that scale, and California decided that frontier models sold into its market should carry a shutoff an outside auditor can verify.