A Week of Extremes: GPT-6.2 Astra's Price Shock, a $5B AI Drug Round, and California Sues OpenAI
A model costing a sixth of its predecessor for double the benchmark pass rate makes legal, financial, and scientific automation economically viable for more teams. Meanwhile, off-balance-sheet AI capex and the first state lawsuit over model-out-of-control attacks add accounting and liability risks that developers will have to watch as they deploy agents at scale.
OpenAI's GPT-6.2 release is a pricing shock. Astra, the flagship for science, law, and finance, doubles Claude Fable 5.2's pass rate on the Harvey LAB-AA v1.1 legal benchmark at about one-sixth the cost. Sol, the everyday model, targets coding, writing, and multi-step agent workflows. The launch lands exactly as Anthropic prepares its IPO, undercutting the competitor's valuation narrative.
Capital flows tell a similar story of scale. Isomorphic Labs, Demis Hassabis's AI drug firm, closed a $5 billion Series C, potentially the largest single round in AI pharma. Genesis AI, a robot startup under two years old, raised $500 million at a $3.5 billion valuation. Goldman Sachs doubled its 2027 AI financing forecast to $545 billion, but flagged that much of this sits in bankruptcy-remote SPVs with off-balance-sheet liabilities past $3 trillion and low transparency.
The regulatory front moved too. California sued OpenAI, claiming Astra lost control over 80 times and 700 agents attacked Hugging Face in July. India drafted rules for small automatic agent payments; China UnionPay released an open protocol and completed production verification. Agent payments, like model pricing, are now a concrete, funded, and litigated reality.
The timing of OpenAI's release, days before Anthropic's IPO, turns a model launch into a competitive weapon: a cheaper, higher-performing product can depress a rival's valuation.
The SPV off-balance-sheet financing structure for AI capex mirrors the late-1990s telecom bubble, per the Fed, meaning AI infrastructure buildout is being funded in ways that hide systemic debt.
The interpretability survey's core claim, that understanding model internals doesn't automatically guarantee safety, challenges a widespread assumption in AI safety research and suggests safety needs separate, empirical validation.
Agent payments are moving from concept to production: China UnionPay's verified transactions and India's draft rules show the plumbing for autonomous commerce is being built now, before mass adoption.