The Product Desk

Google Research and Technion tested 13 models across more than 4 million responses. The models already encode 95-98% of benchmark facts, yet fail to surface 26-34% of them when asked directly. Extra inference-time thinking recovers 40-65% of those misses, and only 10-20% of facts genuinely need reasoning. If your retrieval layer mostly serves stable general facts, it is buying latency and per-call cost rather than accuracy.

Anthropic's Fable 5.1 reads cached context 75% cheaper than Fable 5 — roughly 25% off normal use and up to 45% on agentic loops that re-read the same context every iteration. The model's five-level effort dial ships defaulted to high. Anthropic's own CursorBench numbers show 5.1 at low effort outscoring Fable 5 at high effort for a third of the cost. The saving is a config change someone on your team has to make, and the migration carries breaking parameter changes.

Dharmesh Shah's September 2 essay argues AI collapsed first-order build cost to minutes while leaving maintenance and per-session user cognitive load untouched. Any RICE or value/effort score with build cost in the denominator now reads almost every feature as a quick win. In the same window, Decagon's Jesse Zhang told a16z speedrun that competitive density is a market-size signal rather than a penalty. Two of the three standard inputs in your rubric are pointing the wrong way.

Judge Leonie Brinkema rejected structural remedies in the DOJ's ad tech case, per The Information's reporting, leaving Google's full stack intact. Google's winning argument was that the open web ad market is shrinking: network ads fell from 17% of its ad revenue in 2018 to 9% in Q2 2026, partly because its own AI search sends less traffic to sites. Any acquisition model compounding on Google-sourced organic traffic is compounding on an asset Google just told a federal judge is depreciating.

ASML is raising equipment prices over the objections of TSMC, its largest customer, per The Information's chip roundup — the clearest sign that fab capex baselines are rising, not falling. Google separately contracted 400MW of Fervo geothermal power that does not deliver until 2028, with an option for 600MW more by 2030. Market-implied odds of a Fed rate hike nearly doubled to roughly 70% in a week, per Morning Brew. Model your AI feature margins on flat hardware costs for the next four quarters.