Near-frontier on a laptop, and prices down a quarter

The board

▲ AMD 6.5% ▲ 688825.SS 4.4% ▼ AVGO 5.9% ▼ AMAT 5.1% ▼ NET 4.6% ▼ PATH 4.0%

Markets — Friday's session, the last before the weekend, split the complex rather than moving it: AMD +6.5% and CXMT +4.4% in Shanghai, against Broadcom −5.9%, Applied Materials −5.1%, and the software names giving ground — Cloudflare −4.6%, UiPath −4.0%, CrowdStrike −3.8%, Workday −3.8%.
Open weights — Qwen3.8-27B has taken over Hugging Face trending: 267,725 downloads in thirty days and 10,193 likes, with its own GGUF build trending alongside it.
Models — Opus 5 still holds both ends of the model board: top intelligence at 63.1, best value at $10/M blended.

The pricing consequence arrived in the same window. Prices for models from the leading US labs are down almost a quarter since mid-July on Silicon Data's index, and the Financial Times attributes the cuts to Chinese competition — which moves the driver from lab strategy to competitive necessity, and answers a question we have had open for a fortnight. DeepSeek now charges double at peak against off-peak, with the peak windows mapping onto the Chinese working day. Gemini 3.7 Flash scored 56, which puts Google seventh on intelligence while sitting on the speed frontier. And read against all of it, an argument that the labs are trading world knowledge for reasoning — fewer active parameters, higher reasoning scores, and something quietly given up in between.

Four items complicate that picture. CoreWeave has contracted A100-class silicon out to 2029: either evidence that accelerator generations earn well past the useful lives every depreciation model assumes, or an artefact of scarcity, and the next generation's launch will separate them. Canva's backers wrote down US$7.1bn, with Canva's own mark lower still, after difficulty rolling out AI tools given the cost of frontier models. A grey market in resold inference credits is running at 30–80% off list — what you would expect to see if committed capacity has outrun actual use. And the quarter-end filings had their own tells: Nvidia holding US$21bn of SpaceX, Berkshire adding US$17bn of Alphabet.

Policy moved on four fronts. The White House is preparing to fold open models into its AI framework according to Wired's sources — the existing framework is unpublished and covers closed models only. A draft US letter would ask the 35 signatories of its AI Opportunity Statement to pick a side, though it has not been sent. Taiwan lifted its 2026 GDP growth forecast to 11.05% — a projection rather than an outcome, and an extraordinary one for a developed economy. And Windows OEM licence prices are reportedly up 7–10% on top of memory costs, a claim relayed through three publications back to a Taiwanese report, so hold it loosely.

Then the agents, out in the world and under measurement. Andon Labs' shop-manager agent recommended dismissing a human worker — but only after a nudge to check the attendance policy it had written months earlier and then lost track of. Anthropic's Frontier Red Team found agent swarms fail by all making the same choice, which is a failure mode no single agent exhibits and no amount of redundancy fixes. A verifier found 39.5% of "correct" AI-generated GPU kernels broken once tested against a stricter instrument, while an agent loop produced a 232x kernel speedup, placing 12th of 183 entrants. Both are true at once, and the distance between them is where the engineering now lives. OpenAI's own working paper, across 17 million enterprise messages, is the best data yet on how the tools are actually used, and two opposite readings of the context window as working memory frame what to make of any of it.