2026-09-25政策Appeals court lets the Pentagon exclude Claude, and says guardrails can count as a supply-chain riskA split D.C. Circuit panel ruled that Anthropic's refusal-by-training is 'manipulation' under a 2018 procurement law, whatever the company's motive.阅读约5分钟→
2026-09-24产品Google's Gemini avatars go GA, and the video bills at 6,192 tokens a secondGemini 3.8 Live with Live Avatar is generally available in Gemini Enterprise. Google published the token conversion rates, which makes the cost of visual presence unusually easy to calculate — about $0.37 per minute of speech, roughly 21 times the audio it accompanies.阅读约4分钟→
2026-09-23研究Claude flagged a CRISPR-shaped enzyme system. Nobody knows what it does.Anthropic's new life sciences lab says ~950 Claude agents burned 21 hours and 210 million tokens to surface ART, a reverse transcriptase sitting next to a CRISPR-like repeat array. Its function is unestablished, the underlying enzyme was already known, and humans did every experiment.阅读约5分钟→
2026-09-22模型发布Anthropic ships Claude Opus 5.5, and the headline is the priceThe new Opus lands at $4 and $20 per million input and output tokens, down from $5 and $25, with Anthropic claiming Fable 5.1-level results on most work. It is also the company's first model since Dario Amodei argued publicly for pacing the frontier.阅读约6分钟→
2026-09-21研究OpenAI's new math advisory group can say anything. It can't slow anything down.Nine mathematicians, hosted at the Institute for Advanced Study, can assess OpenAI's results and go public with objections. The charter explicitly excludes advising the company on how fast to move — the exact complaint twenty-five Fields Medalists made ten days ago.阅读约4分钟→
2026-09-18政策Anthropic's first embedded evaluator is Accenture, and Anthropic is paying for itEach side commits at least $1 billion over five years to put outside evaluators inside Anthropic with employee-level access. The evaluator is a consultancy that already sells Claude to enterprises.阅读约4分钟→
2026-09-17政策Anthropic says Claude now leads 26% of its own AI researchAnthropic published three measurements of how AI gets built inside its own walls: the share of its AI research Claude leads, how its 30,000-agent fleet is monitored, and how much compute goes to safety. The figures are the company's own, and nobody outside has audited them.阅读约6分钟→
2026-09-15模型发布Gemini 3.8 Live is generally available, and the Live API's defaults moved with itGoogle shipped two audio-to-audio models to GA on Tuesday. The benchmark table is Google's own; the migration notes are the part that will cost you an afternoon.阅读约6分钟→
2026-09-14政策Trump calls the slowdown a hoax. Microsoft writes down its red lines.The president phoned into a conference on Jensen Huang's speakerphone to call the industry's safety turn a setup. Hours earlier, Microsoft's AI division had published a code of conduct that, as TechCrunch describes it, outranks anything a user or a task asks of its models.阅读约7分钟→
2026-09-10安全Anthropic names seven Chinese labs, and says two served Claude to their own usersA threat report published today puts five distillation campaigns at roughly 190 million exchanges and alleges Moonshot and DeepSeek quietly relayed customer requests to Claude. The attribution is Anthropic's own, and the telemetry is not public.阅读约7分钟→
2026-09-09产品Apple's rebuilt Siri finally has a date. It also has a meter.iOS 27 lands on 14 September with Siri AI in beta — English only, absent from mainland China and from the EU on iPhone, and with daily caps on the cloud half that Apple says will cost money later.阅读约6分钟→
2026-09-08研究OpenAI's Navier–Stokes claim comes with a Lean proof. Its provenance is contested.OpenAI published a writeup and a Lean formalization for a finite-time blowup result on one of the seven Millennium Prize problems. Hours earlier, an NYU mathematician had accused it of building on his unpublished work.阅读约6分钟→