The week in three numbers: $2 / $10 Sonnet 5's intro price per million tokens (near-Opus-4.8 agents) · Aug 2 the EU AI Act's GPAI enforcement powers switch on (€15M / 3% fines) · 3 weeks from Fable 5 gated to Fable 5 global again.

Five decision-relevant moves this week, tagged by lane.

Last week set three live threads: the economics arrived, the government became a gate, and talent was flowing to Anthropic. This week all three moved — and two reversed direction. Here's the increment.

Last week's lesson was "route routine work to the cheapest model that clears the bar," and the cheapest models were open weights (DeepSeek). This week Anthropic contested that directly: Sonnet 5, a near-frontier agentic model, at an intro $2 / $10 and set as the default. Google pushed the same way in media (Nano Banana 2 Lite at ~$0.034 an image). The majors aren't conceding the low end — they're pricing into it, because losing the routine-workload tier to open weights is an existential threat to the API business the IPOs are valued on. It's Ethan Mollick's demand-side thesis — chatbot → agents that do hours of unattended knowledge work — arriving with a price tag, which is exactly what makes the per-workload routing question urgent.

The lesson: the model-portfolio discipline from Issue 002 is still right, but the inputs are unstable. "Cheapest capable model" is now something the incumbents re-contest every few weeks — so treat routing as a standing function with a live price feed, not a one-time decision. Don't hard-wire your stack to last quarter's price sheet.

Last week the novel development was the US gating frontier releases. This week the gate came off: the Fable 5 / Mythos 5 restrictions (imposed June 12 after researchers found a jailbreak that bypassed Fable 5's safeguards) lifted the moment Anthropic shipped a classifier that blocked it — Fable 5 is global again as of July 1; Mythos 5 is back but reportedly still limited. And the mechanism underneath — the June 2 EO — turns out to be a voluntary cyber review with preclearance banned, not a statutory chokepoint. This also closes part of last week's "safety-as-security" watch: a capability was paused and then released on a concrete, shippable safeguard — but the whole exchange was about cyber/jailbreak risk, still not alignment.

The lesson: US model availability is now a fast-moving, capability-triggered variable — a model can vanish on a red-team finding and return on a patch, in weeks. That's operationally very different from the EU's statutory Aug 2 regime. For procurement and risk, plan for availability volatility in US frontier models, and watch the EU date for what durable governance actually costs.

The John Jumper hire was three weeks ago; this week it's a flagship. Claude Science — a research workbench (Claude Code, but for scientists) wiring 60+ scientific databases and computation tools into one workspace, available to all paid subscribers — turns a Nobel-laureate acquisition into a shipping product with a revenue story, and, notably, into Anthropic pursuing its own drugs for rare, neglected diseases. The talent flow we've tracked since Issue 001 isn't just prestige signaling; at Anthropic it's being metabolized into product lines on a pre-IPO clock. (STAT, MIT Tech Review)

The lesson: watch where talent lands and how fast it ships. The speed of Jumper → Claude Science is the tell: Anthropic is converting hires into product lines faster than a public company typically can — which is both the bull case for its IPO and the reason its safety review is running against a stopwatch.

Last week's read was that cost discipline beats model selection, and the gate was up. Both softened this week — and that complicates the story rather than refuting it. The position holds: "measure, then route" still beats "standardize on one" — but it now requires a live subscription to price and availability, because the incumbents will keep re-cutting prices and the US gate will keep swinging on safety findings. The honest counter-case: one intro-priced Sonnet doesn't erase a still-large gap to the cheapest open weights (DeepSeek's mid-July V4 could reopen it within the month); a voluntary gate that lifts in three weeks is ad-hoc governance, not a durable regime — the EU's statutory Aug 2 is the real test of whether rules bite; and "talent → product in three weeks" is exactly the velocity that outruns careful safety review. All true, which is why the operator move is flexibility, not conviction in any one vendor or any one week's price.

1) Does DeepSeek's mid-July official V4 (peak/off-peak pricing) reopen the price gap Sonnet 5 just narrowed? 2) Does the EU AI Act's Aug 2 enforcement actually bite — the statutory counterpart to the US's reversible gate? 3) Does Google answer its talent bleed with a hire or a retention, or keep answering with launches? 4) Does Claude Science produce a validated result — a real candidate, a replicated finding — or stay a productivity layer?

Not who-said-what — the live split between serious people, and where we come down.

Only reads that add something the front didn't already give you.

What's coming — dated anchors worth calendaring.

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

Issue 003 note: assembled from comprehensive public reporting for June 29 – July 5, 2026, and built incrementally on Issue 002. 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 a fixed spine of the field's highest-signal voices — operators, researchers, and independent skeptics like Andrew Ng, Ethan Mollick, Simon Willison, Nathan Lambert, the AI Snake Oil team, Erik Brynjolfsson, Cassie Kozyrkov, and Eric Topol — plus a rotating edge of specialists (Chip Huyen, Jack Clark, Ben Thompson, Robert Wachter, and others) and the primary research, regulator, and lab feeds. We tag every source — vendor, researcher, operator, investor, regulator, economist, or skeptic — and check strong claims across categories, so we curate evidence, implementation, and disagreement rather than celebrity.

The brief comes in two speeds: a fast skim — the single most important development, three concrete moves, and the week's decision-relevant signals — then a longer Synthesis that connects them, 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."