Benchmark-record week. Gemini Deep Think, GPT-5.5-Cyber, and GLM-5.2 each reportedly topped some leaderboard. Treat single-vendor records as marketing until someone replicates them. This week the decision-relevant variables weren't the scores — they were access and cost: who's allowed to use a model (GPT-5.6's government gate), and what it costs to run (the DeepSeek price crash). Read the access model and the price sheet, not the leaderboard.

Issue 001 opened three threads. This week two of them escalated, a fourth broke open, and one item needs a correction. Here's the increment.

Last week's "watch where decorated people go" is no longer ambiguous. It's a flow out of Google DeepMind into Anthropic — five senior researchers in six days, concentrated in pre-training and coding — and this week it widened beyond research, with Apple's Vision Pro chief reportedly going to OpenAI. The reported driver is pre-IPO equity at OpenAI and Anthropic.

The lesson (updated): the Issue 001 signal held and broadened. If you're betting on a lab's trajectory, weight Anthropic's pull on frontier talent — and watch whether Google answers with retention or a marquee release (it tried the latter this week with Deep Think).

For a year the incentive was to use as much AI as possible. That just inverted. Open-weight models got cheap enough — DeepSeek reportedly a fraction of frontier cost — that enterprises are instrumenting token spend, imposing per-engineer caps (Uber), and routing routine work to the cheapest model that clears the bar (Microsoft, Lindy). And for the first time there's hard labor-market evidence attached: Brynjolfsson's live dashboard shows AI compressing entry-level hiring in exposed roles. This is the first thread in two issues aimed squarely at the CFO and the CHRO, not the CTO.

The lesson: the frontier premium now has to be justified per workload. Run a model portfolio (cheap/open by default, frontier reserved for the hardest tasks), instrument spend, and US-host any open-weight model for security. And start asking the workforce question now — if AI eats the junior tier, where does your senior talent come from in five years?

The most novel development of the week: the US gated a frontier release (GPT-5.6) for the first time, OpenAI objected publicly, and the rest of the board moved in sympathy — Anthropic's export ban hit week 3 and spawned Asian "Mythos-like" replacements, Anthropic escalated its distillation fight to name Alibaba, and China's Zhipu hit a record valuation as it closes the gap. Model access is now a geopolitical variable, not just a pricing one.

The lesson: provenance and sovereignty are entering procurement. Where a model is hosted, who can legally use it, and whose weights they are now matter for compliance and risk — exactly as the cheap-open-model wave (Thread 2) pulls teams toward Chinese weights. Those two forces are on a collision course; "US-hosted open model" is the seam between them.

Last week we asked: will a lab pair a capability release with a concrete safeguard? This week gave a partial answer: OpenAI's Daybreak (GPT-5.5-Cyber + named partners) is the closest yet — but it's cyber-defense, not the recursive-self-improvement brake Anthropic argued for. In parallel, Dario Amodei's recent "Policy on the AI Exponential" essay dropped Anthropic's transparency-only stance for mandatory third-party testing and government power to block deployments — a notable hardening. And a sober counterweight from healthcare: a peer-reviewed RCT found automation bias persists even in AI-trained physicians (accuracy fell when the AI was wrong), undercutting "human-in-the-loop" as a sufficient safeguard. (Amodei, NEJM AI)

The lesson: the labs' public safety story is consolidating around security — concrete, fundable, revenue-adjacent. Watch whether it generalizes to the harder alignment questions, or stops at the ones with a product attached.

For an executive, the week's signal is not a model — it's that AI got cheap enough that cost strategy beats model selection, and consequential enough that its labor and geopolitical bills are now arriving. The leaderboard is the most-marketed, least-useful artifact in the field; the operators who win this quarter will instrument spend, run a model portfolio, and start asking the workforce question. The honest counter-case: frontier models still lead the hardest reasoning, cheap open-weight models carry real reliability, governance, and provenance risk (especially Chinese weights), and "AI eating the on-ramp" is early, single-source data that serious economists dispute. All true — which is why "measure, then route" beats "standardize on one," in both your spend and your hiring.

1) Do the open-weight pilots survive Q3 once reliability and governance bills land? 2) Does the EU AI Act's Aug 2 enforcement actually bite, given many states/members haven't named authorities? 3) Does Google answer the Anthropic talent pull? 4) Does the safety-as-security pattern reach alignment, not just cyber?

Got a mind we should be reading, or a correction? Reply and tell us.

Issue 002 note: assembled from comprehensive public reporting for June 22–28, 2026, and built incrementally on Issue 001. Items marked "reported" are as-reported and not independently confirmed; we link primary sources where they exist and hedge where they don't.

AI Above the Cut is a weekly brief for executives — VP-and-up leaders in strategy, healthcare, and AI transformation who want signal over noise. Each Sunday we read the week's output from a fixed roster of the field's highest-quality thinkers — researchers, lab leaders, and operators like Andrew Ng, Andrej Karpathy, Fei-Fei Li, Demis Hassabis, Dario Amodei, Ethan Mollick, Cassie Kozyrkov, and Erik Brynjolfsson — plus the primary research, regulator, and lab feeds.

The brief comes in two speeds: a 4-minute Quick Hits skim, then a longer Synthesis that connects the week's moves, takes a position, and links to the primary work. We optimize for quality over influence, link to the source (the paper, the post, the talk) rather than the hype around it, and flag anything unconfirmed. No "10 AI tools you need today."