Clinical rankings, scientific evidence, agent delegation, regional access, and safety tests all changed when the surrounding system changed. The week's leverage shifted toward task distributions, external validators, operating-system permissions, cloud routes, and adversarial test design, leaving the model checkpoint as only one input to performance and availability.
1. Clinical AI rankings reverse when evaluation moves closer to care
AMIE received appropriate-plan ratings of 95%-98%, against 72%-81% for 21 physicians, across 100 constructed multivisit cases. BRIDGE then placed the leading tested model at 44.8% on 87 tasks drawn from real clinical text despite medical-exam performance as high as 92, while SpeechDx found no audio representation that transferred reliably across 27 clinical-speech tasks.
The findings conflict only if "medical capability" is treated as one property. AMIE supplied a text-only simulation, retrieved guidelines, and cases with definitive answers; BRIDGE added heterogeneous records, languages, specialties, and coding work, while SpeechDx changed the modality entirely. A high score can clear one measured workflow, but it cannot carry authorization from simulated management into records, billing, or clinical speech.
Sources: Nature study of AMIE · Nature Biomedical Engineering study of BRIDGE · SpeechDx preprint
2. External validators turn fluent scientific work into measurable progress
GPT-Rosalind passed 36.1% of LifeSciBench overall and 71.1% of its nine scientific-communication tasks, yet only 14.8% of numeric tasks; adding artifacts or URLs cut its pass rate from 45.1% to 28.1%. In chemistry, GPT-5.4 and Maria ran 10,080 reactions, lifted mean yield from 16.6% to 25.2%, and had human chemists reproduce improvements in 11 of 14 substrate pairs.
Formal structure alone supplied no equivalent check: a ContractNLI study found the highest benchmark accuracy from model-generated formal reasoning even as models returned classifications inconsistent with the Z3 solver. The leverage comes from an external state that prose cannot negotiate, whether a calculation, solver output, assay, or bench-scale reaction. Research throughput can rise before trust in final answers does, shifting the scarce capacity to experiments and expert verification.
Sources: OpenAI's LifeSciBench report · OpenAI's AI-chemist study · Legal-reasoning preprint
3. Agents amplify specification quality rather than eliminating it
Across roughly 400,000 Claude Code sessions, people made about 70% of planning decisions while Claude made about 80% of execution decisions; expert-rated prompts triggered 12 actions on average, versus five for novice-rated prompts. Hierarchical subgoals separately raised vague web-task pass rates from 76.7% to 90.7%, but added nothing when instructions were already precise.
Android 17 extends the same division into products by exposing declared app functions and local state as agent tools. Project Fetch showed its physical limit: Opus 4.7 completed five robot integration tasks in about 12 minutes yet failed the final closed-loop ball-control task. Implementation speed compounds good specifications, while ambiguous goals, permission scope, and feedback-sensitive control remain the points where delegated work loses leverage.
Sources: Anthropic's Claude Code study · Hierarchical-demonstration preprint · Android 17 developer announcement · Anthropic's Project Fetch report
4. Sovereign AI fragments across capital, cloud routes, and employee geography
Sarvam announced a $234 million first close of a planned $300 million Series B at a $1.5 billion post-money valuation, led by HCLTech and Bessemer Venture Partners, to expand domestic models and infrastructure. In the same week, Microsoft reportedly extended OpenAI models into China through its cloud business, while JPMorgan cut Anthropic access for Hong Kong staff even as colleagues elsewhere retained it.
These moves pull sovereignty in opposite directions: ownership and local training capacity concentrate control, but resellers can project a foreign model across borders and provider policy can remove it office by office. Model weights, contracting entity, inference location, data route, and user entitlement can each sit in a different jurisdiction. One multinational can therefore run different AI stacks behind the same internal application solely because an employee crosses a geographic boundary.
Sources: Sarvam's Series B announcement · Bloomberg on Microsoft's China model sales · Financial Times on JPMorgan's Hong Kong restriction
5. Common-case simulations and adaptive attacks measure different safety boundaries
OpenAI replayed about 1.3 million de-identified conversation prefixes with candidate models and achieved a median 1.5-fold error when forecasting undesirable-behavior rates after release. The method cannot estimate events rarer than roughly one in 200,000 messages. Separately, an automated red team found static obfuscation almost fully blocked but adaptive attacks succeeding on 11.5% of tested intents against Opus 4.8 and 6.1% against Fable 5.
The apparent tension comes from sampling different distributions. Deployment replay estimates recurring behavior under representative traffic; iterative search concentrates attempts on a boundary until it fails. The former can calibrate expected incident volume while the latter estimates attacker leverage, and neither substitutes for the other. A release can show low predicted prevalence and remain cheaply breakable under sustained attack, creating two separate gates for routine deployment and hostile exposure.
Sources: OpenAI's deployment-simulation study · Automated red-team preprint