Amodei calls for an AI slowdown, and Anthropic accuses Chinese labs of routing 35 million requests to Claude
Plus: Mistral lands inside Firefox, Gemini 3.8 Live ships, and six new models arrive in seven days.
The week on the beat
1. Dario Amodei called for an AI slowdown, and the argument swallowed the week
Altman and Musk backed a version of it, Microsoft set limits on future frontier models, and Anthropic's three-step "pace the frontier" plan drew support from OpenAI, xAI and Microsoft. Trump rejected it outright and attacked Amodei by name, China's state press called it a Cold War tactic, and Congress started arguing about safety-testing language with no bill close to passing. Over a hundred outlets covered it in six days, which is the most concentrated coverage of anything this year.
2. Anthropic says seven China-based labs ran industrial-scale distillation against Claude
The allegation is that Kimi users' requests were served by Claude and returned as Moonshot's own output, at least 35 million times by Business Insider's count, with Alibaba and DeepSeek named in the same report. Leave the ethics argument aside. The practical point for anyone building on a Chinese frontier model is that the capability you benchmarked may partly belong to a vendor who can revoke it, and Anthropic has now shown it will go public and cut routes.
3. Mistral's models now run inside Firefox, on the user's own machine
Mozilla's new Smart Window assistant uses Mistral for local translation and text processing, executing on the device rather than in the cloud, which makes it private and offline by default. This is the first time a major browser has shipped a named third-party model as a local feature, and it gives Mistral consumer distribution it could not buy.
4. Google shipped Gemini 3.8 Live and 3.8 Live Extended Thinking
Extended Thinking spends more compute planning before it answers, aimed at multi-step and technical work, and Google is positioning Live directly against OpenAI's GPT-Live-1 on price. Neither variant has reached our registry yet, so treat the pricing claims as Google's until they show up in a public tracker.
5. Anthropic folded Claude Chat and Cowork into a single product
The merged interface decides on its own whether a request needs a straight answer or a longer agentic workflow, so the routing choice moves from the user to the model. Claude Docs and Claude Slides arrive alongside it. If you were building on the split between the two surfaces, that distinction is gone.
6. Sakana released Fugu Max and Fugu Ultra v2 for multi-agent orchestration
The pitch is a better cost-to-quality trade-off when several specialised agents work together, rather than a single stronger model. Fugu Max lists at $2 and $6 per million with a 1M context, Ultra v2 at $5 and $30.
Model moves
- No price changes this week, despite appearances. Two moves looked real and neither was. DeepSeek V4.1 Flash reads as a 50 percent cut because DeepSeek prices it on a peak/off-peak schedule where peak is exactly double, so a tracker reports a different number depending on the hour it checked. Tencent's Hy3 fell 38 percent on September 11 and returned to exactly its old price on September 15, which is a four-day sale. Both were checked against the vendors' own pricing pages.
- No benchmark movement at all this week. No tracked model's reported score changed, and no suite was re-run. The last bulk revision was Epoch's SimpleQA Verified re-score across 23 models on August 28.
- New: Fugu Max (Sakana). $2/$6 per 1M with a 1M context. Model page →
- New: Fugu Ultra v2 (Sakana). $5/$30 per 1M with a 1M context. Model page →
- New: DeepSeek V4.1 Flash (DeepSeek). $0.15/$0.60 per 1M off peak and $0.30/$1.20 at peak, with a 1M context. Peak is 01:00 to 04:00 and 06:00 to 10:00 UTC on weekdays. Model page →
- New: Ling 3.0 Flash VL (inclusionAI). $0.06/$0.18 per 1M with a 131K context, and vision. Model page →
- New: Schematron V2 Small (Inference.net). $0.05/$0.23 per 1M with a 128K context. A Turbo tier landed the same day at $0.03/$0.15. Model page →
Personal take
Even as the industry discussed slowing down this week, six new models still launched. Sakana released two, Inference.net added two more, inclusionAI launched one, and DeepSeek's V4.1 Flash came out on Monday. If there really was a slowdown, you would notice it through new access limits and stricter terms, which is what's happening now at Anthropic and Moonshot. You won't get a warning email if things pause; instead, your usual access might just disappear. This week, it's smart to have a backup for any model you can't run on your own.
Until next Thursday, Anmol
Get the next issue
Free, weekly, unsubscribe anytime. That’s the whole pitch.
Free forever. No spam. One-click unsubscribe. See our Privacy Policy.