The AI Brief

OpenAI CEO Sam Altman published the framework on September 21 and will brief an open meeting of the UN Security Council's 15 members in person on Wednesday, as world leaders gather for the UN General Assembly's high-level week, which runs September 22 through 28. The framework calls on the US to utilize existing AI safety organizations, including the Center for AI Standards and Innovation (CAISI), and to work with allied nations to build shared technical benchmarks for capability measurement, risk assessment, and safeguard sufficiency.

On RSI specifically, the proposal outlines three proposed standard categories: evaluation criteria for measuring how much autonomous AI research is occurring inside a given lab; human-oversight rules specifying which categories of automated research processes must trigger immediate human review; and a common incident-classification and reporting framework for alignment failures and automated-research anomalies. OpenAI cited its own misalignment reporting framework and its research acceleration report, both published in September, as early contributions to the effort. The proposal states that "fully autonomous RSI is not happening today, and we should not pursue it unless and until it can be done safely."

The push arrives at a moment of friction with the Trump administration, which has framed global AI governance primarily as a competitive race against China rather than a safety-first exercise. OpenAI's framing attempts to bridge the gap: it argues the US is "well positioned to lead because its AI industry is at the technical frontier," appealing to competitive logic rather than safety warnings. The Security Council operates by consensus among five permanent members, meaning any binding resolution would require Chinese and Russian agreement; Monday's proposal is shaped accordingly, centering voluntary technical standards rather than treaty obligations.

The New York Times reported on September 18 that Anthropic is on pace to generate more than $100 billion in annualized revenue this year, citing people familiar with the matter. The company's run rate stood at roughly $9 billion at the end of 2025 before rising to $47 billion in May and more than $65 billion by end of July, per Bloomberg. The acceleration is primarily attributed to enterprise adoption of Claude Code and the Cowork platform. Anthropic's investors expect the company to finish 2026 between $100 billion and $120 billion in annualized revenue, the Financial Times reported. Reuters places the IPO roadshow starting mid-October at the earliest, with a Nasdaq listing potentially in November. A $2 trillion IPO valuation has been cited by multiple outlets citing sources, though Anthropic has not confirmed that figure. OpenAI, by comparison, doubled its revenue to roughly $40 billion annualized by mid-2026.

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

SpaceXAI disclosed that Grok 4.7's training incorporated supplemental data from SpaceX operations: Starlink satellite telemetry, manufacturing records, and engineering failure logs. The stated goal is a model that reasons better about hardware and physical systems than models trained solely on public internet text. On EEBench, an electrical engineering benchmark, Grok 4.7 scored 64.0% versus Fable 5.1's 56.4% and GPT-5.6 Sol's 39.4%, the largest performance gap in any benchmark category tested. On the Harvey Legal Agent Benchmark, Grok 4.7 scored 19.6% versus Fable 5.1's 6.7% and GPT-5.6 Sol's 2.5%, a domain where SpaceX's legal and regulatory corpus is plausibly large. The pattern suggests the SpaceX data is producing measurable gains in structured technical domains rather than general knowledge.

The structural implication is significant: xAI, now part of SpaceX, has access to operational data generated continuously by one of the largest satellite constellations and a high-cadence rocket program. That data is not available to OpenAI, Anthropic, or Google, and cannot be synthesized from public sources. If training-time operational data produces reliable benchmark lifts, the competitive advantage accrues to platforms with live operational systems, rather than to labs that can buy the most GPU time. Whether the advantage persists into the next training generation remains to be seen; SpaceXAI released no information about what fraction of training compute the SpaceX data represented.

SpaceXAI, the AI division now operating inside SpaceX following the xAI and SpaceX merger, launched Grok 4.7 on September 21, 2026, at 16:17 UTC. The model carries 2.1 trillion parameters, a 40% increase from Grok 4.6's 1.5 trillion, and a 500,000-token context window. It is available through the xAI API, Grok Build, Cursor, and GitHub Copilot (rolling out across Pro, Pro+, Max, Business, and Enterprise plans). Pricing is unchanged from Grok 4.6: $2 per million input tokens and $6 per million output tokens for prompts under 200,000 tokens, doubling above that threshold.

On self-reported benchmarks: CursorBench 4.0 (long-horizon software engineering) 46.3%, ahead of GPT-5.6 Sol's 41.7% but behind Fable 5.1's 51.8%; DeepSWE v1.1 at high effort 71.0%, between Sol's 72.7% and Fable 5.1's 70.0%; EEBench (electrical engineering) 64.0%, leading all tested models; Harvey Legal Agent Benchmark 19.6%, also leading; Terminal-Bench 4.0 38.0%, well behind Fable 5.1's 57.9%; HealthBench Professional 56.7%, behind both Sol and Fable 5.1. On the GDPval knowledge-work ranking (an Elo-style score across professional tasks), Grok 4.7 posted 1695 versus Fable 5.1's 1735. On AA Briefcase v1.1, a multi-hour office-task benchmark, it scored 1657 Elo, also behind Fable 5.1's 1678.

Elon Musk had publicly committed to a Grok 4.7 release at least five times since late July, giving dates that slipped progressively from "four weeks out" in late July to "needs a few more days to cook" on September 11, making the final release 31 days past the first announced target. The delay was attributed to the SpaceX engineering data integration. SpaceXAI also disclosed that invite-only red-team access for Grok 4.7's cybersecurity capabilities is available to selected security research partners.

SB 947, the No Robo Bosses Act of 2026, authored by California State Senator Jerry McNerney (D-Pleasanton), passed the Assembly 53-14 and the Senate 28-10 before reaching Newsom's desk on August 31. The September 30 deadline applies under California law. The bill bars employers from relying solely on an automated decision system (ADS) to terminate or discipline workers, and requires meaningful independent human verification when an ADS informs such a decision. It also requires employers to notify workers after a disciplinary decision was made using algorithmic input. First covered in Vol. I, No. 108 (September 6, 2026).

Newsom vetoed the predecessor bill, SB 7, in October 2025, citing overbreadth and overlap with existing employment law. The 2026 version removes the blanket proactive notification requirement that Newsom found objectionable and narrows the scope to termination and discipline specifically. The California Chamber of Commerce removed SB 947 from its "Cost Drivers" list after the revision, a notable shift. Whether Newsom views the changes as substantive or cosmetic is the operative uncertainty. If he signs, violations expose employers to administrative penalties and civil liability under existing Fair Employment and Housing Act frameworks. If he vetoes again, the operative law remains the status quo.