The AI Brief

SB 947, authored by California State Senator Jerry McNerney (D-Pleasanton), passed the Assembly 53-14 and the Senate 28-10 before reaching Newsom's desk. Under the statute, employers that rely primarily on ADS output to make termination or disciplinary decisions must have a human reviewer corroborate that decision using additional information such as managerial evaluations, peer reviews, and personnel files. Workers subject to those decisions must receive written notice that AI was in the loop, a description of the employee data the system drew on, and the name of a human point of contact who can explain the decision. Violations carry a civil penalty of $500 per violation, with punitive damages and attorney fees potentially applicable.

Newsom vetoed a predecessor bill, SB 7, in October 2025, writing that it proposed overly broad restrictions. The 2026 version narrowed the scope: where SB 7 barred ADS use in termination decisions broadly, SB 947 targets reliance primarily on ADS output and requires corroboration rather than prohibition. Newsom's reversal reflects both the narrowing and the sustained political pressure from the California Federation of Labor Unions, AFL-CIO, which sponsored the bill. The law takes effect July 1, 2027. First covered in Vol. I, No. 128 (the prior edition noted that Newsom's inaction would convert the bill to law without a signature; he in fact signed it actively, a materially distinct development).

Enforcement will operate through California's civil court system rather than a dedicated AI agency. The operative question for multistate employers is compliance readiness: performance-management software sold into California will need a human-corroboration layer and an audit trail showing that AI recommendations were verified, not simply approved, before action was taken.

In a joint announcement with Anthropic on October 1, Barclays said it is expanding its strategic collaboration to integrate Claude across three distinct workloads: software development and legacy system modernization, an internal Colleague Knowledge Assistant, and email routing inside Global Markets. The Knowledge Assistant, live since 2025 and built on a retrieval-augmented generation (RAG) architecture, has been adopted by more than 16,000 colleagues and handled over one million searches. The Global Markets email-triage platform processes approximately 120,000 client-inquiry emails per day, with Claude models classifying incoming requests, enriching them with relevant information, and routing them to the appropriate workflow. Both the 16,000-colleague and 120,000-email figures are reported as current operational data; the 50% developer adoption target for Claude Code by year-end and a majority-of-software-engineers target for 2027 are forward-looking projections. The announcement was jointly made by Barclays Group Co-Chief Operating Officers Anne Marie Darling and Craig Bright, a dual-COO sponsorship that signals operational commitment rather than pilot status. Anthropic Chief Commercial Officer Paul Smith was also quoted in the joint statement.

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

Google confirmed the launch on its research blog on September 24, with the post authored by Travis Beals, Senior Director, Paradigms of Intelligence. The MVP satellite, roughly the size of a household refrigerator, was built by Planet Labs and carries four Trillium TPUs fed by solar arrays rated at approximately one kilowatt of power, enough to run the chips in 15-minute compute bursts before powering down for thermal cooling. Google's target Gemini workload for the satellite is short inference queries processed during those windows. The mission is expected to operate for up to one year in dawn-dusk sun-synchronous low Earth orbit.

The full Suncatcher architecture envisions clusters of 81 satellites flying within a one-kilometer radius at around 650 kilometers altitude, linked by free-space laser interconnects running at 1.6 Tbps. Google and Planet are planning a two-satellite mission in 2027 to test those laser links, which are the load-bearing technical dependency the single-satellite launch cannot evaluate. Google Senior Vice President for Research James Manyika has said publicly that he does not expect anything operationally useful for several years. Analysts covering the orbital compute space, including those at Futurum and independent forecasters, note that cost parity with terrestrial infrastructure requires launch costs to fall from current levels near $1,500 to $2,900 per kilogram toward approximately $200 per kilogram. First covered in Vol. I, No. 125 (September 25) as a confirmed pre-launch announcement; today marks the actual launch event.

The D.C. Circuit's 2-1 decision, issued September 25, upheld the Pentagon's March 2026 designation of Anthropic as a national-security supply-chain risk under FASCA. The majority, written by Circuit Judge Gregory Kastas, found that the Department of Defense "reasonably feared that Anthropic might manipulate Claude's design to prevent it from performing national-security functions that the Department deems contractually authorized and necessary." Judge Karen LeCraft Henderson dissented sharply, writing that enforcing upfront ethical standards does not constitute a supply-chain security threat.

The underlying dispute traces to a $200 million contract signed in July 2025 to bring Claude models to the Pentagon's GenAI.mil platform. Negotiations broke down when Anthropic insisted on retaining restrictions prohibiting fully autonomous lethal weapons use and mass domestic surveillance of US citizens, and reserving the right to trigger a kill switch if it detected out-of-scope use. Defense Secretary Pete Hegseth designated Anthropic a supply-chain risk in March 2026 after those terms failed. Anthropic sued in federal court; U.S. District Judge Rita Lin, presiding in San Francisco, initially issued an order pausing the ban, but the D.C. Circuit's appellate ruling now gives the Pentagon legal backing. A conflicting district-court ruling from Judge Lin and potential further appeals leave the long-term status of Anthropic's government work in legal limbo.

The ruling's broader effect is that any AI company with non-negotiable model-level use restrictions now has a documented legal pathway by which those restrictions can be characterized as a supply-chain risk under FASCA, regardless of whether the company believes those restrictions are safety-essential. Companies supplying AI to defense contractors, not only to the Pentagon directly, face the same exposure because the FASCA designation reaches the supply chain.

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

Anthropic released Claude Sonnet 5.5 on September 28, 2026, six days after Claude Opus 5.5 opened the 5.5 family. The model carries the API ID claude-sonnet-5-5 and is available on the Claude API, Amazon Bedrock, Google Cloud, and Microsoft Foundry. Context window is one million tokens; maximum output is 128,000 tokens (300,000 on the Batch API in beta). Pricing is $2 per million input tokens and $10 per million output tokens, unchanged from Sonnet 5, with cache reads at $0.20 and Batch API at 50% off. The model scored 70.6% on Terminal-Bench 4.0. Adaptive thinking is on by default at high effort. Knowledge cutoff is June 2026.

The 5.5 family completes Anthropic's current product tiers: Opus 5.5 (September 22) is the flagship, Sonnet 5.5 is the mid-tier everyday model, and Haiku 5.5 remains forthcoming. The speed and cost-per-task improvements Anthropic claims come without a per-token price reduction, which means the gains are efficiency-side rather than list-price reductions visible in vendor comparison tables. Teams migrating from Sonnet 5 should review Anthropic's migration notes for breaking changes before switching production routes.

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