In partnership with
Welcome, AI enthusiasts
Anthropic published eight months of evidence yesterday on peoples trying to misuse Claude. Seven Chinese labs show up in it, next to Russian propaganda desks, Iranian state institutions and a French hacktivist who built a doxxing platform on his own. Let's dive in!
In today’s insights:
Anthropic Exposed Chinese AI copying Claude 151M Times
OpenAI drops GPT-Live-1 Closest Model to Natural Voice
DeepSeek's drops 33x Cheaper AI Beats Opus 5
Read time: 4 minutes
LATEST DEVELOPMENTS
Source; The Economist
Evolving AI: Anthropic published eight months of proof on how its own model was being misused.
Key Points:
Moonshot quietly forwarded almost 300,000 customer requests to Claude over ten days instead of running them on Kimi.
DeepSeek moved more than 12 million user exchanges onto Claude across fourteen days in July without telling its customers.
Xiaomi and SenseTime were two of the seven China-based labs Anthropic named, and their traffic carried sensitive information from real users and big companies.
Details:
Anthropic traced 151 million exchanges between May and July 2026 to Alibaba, which pulled reasoning out of Claude Opus 4.6 and 4.7 and fed it into Qwen 3.5, 3.6 and 3.7. The same report covers Russian and Iranian propaganda that ran on radio and television, plus five biological cases Anthropic blocked before establishing the purpose. Each case study names the domains and account handles that sat behind every disrupted operation.
Why It Matters:
Chinese AI labs didn't achieve hyper-efficiency through sheer hard work, but by secretly routing millions of live user queries into Claude to clone its logic. By turning customers into unwitting data mules, companies like DeepSeek and Alibaba triggered a massive privacy scandal and anchored their models to a Western competitor’s worldview. This structural theft kills the open-API era. Western labs will build fortified digital fortresses with strict identity checks to block distillation bots. AI is no longer just software. It is a restricted, national security asset operating behind heavily guarded digital borders.
ChatGPT gives you generic answers because you give it generic prompts.
You know the fix: longer prompts, more context, clearer constraints. But typing all that takes five minutes per prompt, so you shortcut it. Every time.
Wispr Flow lets you speak your prompts instead of typing them. Talk through your thinking naturally — include context, constraints, examples — and get clean text ready to paste. No filler words. No cleanup.
Works inside ChatGPT, Claude, Cursor, Windsurf, and every other AI tool. System-level, so there's nothing to install per app. Tap and talk.
Millions of users worldwide. Teams at OpenAI, Vercel, and Clay use Flow daily. Free on Mac, Windows, and iPhone.
Evolving AI: GPT-Live-1 listens and speaks at the same time, and it is live in the API today.
Key Points:
In early tests at language app Speak, students who stopped to think got interrupted almost 80 percent less.
Full Duplex Bench put it 30 points above GPT-Realtime-2.1, the model OpenAI has been selling until now.
Developers pay 5 cents a minute for the voice layer and choose their own backend model like GPT-6 Astra.
Details:
OpenAI put the listening and the speaking inside a single model, so audio no longer gets handed between separate systems that each cost time. Yelp already runs restaurant calls on it, callers can book a table or order food while noise and side chatter keep going around them. Twelve new voices came with the release and OpenAI added phone line support so an agent can pick up a call.
Why It Matters:
ChatGPT had it first, and GPT-Realtime-2.1 made the caller wait until the model stopped talking. Call centers held out longer than most industries because a voice that cuts people off drives them away. Any company still selling the turn-taking version now has a rewrite ahead of it, and will lose customers to rivals who sound human on the phone. Expect more of your calls to be answered by ai voice models this year, with fewer people sitting behind the line.
👀 Our Tip
There is a good chance a Fable 5.1 Build Day is landing in your city over the next two weeks. Community members organise each one, not Anthropic, and they run from today through September 25. Bring a problem you want to solve or an idea you have not started yet, then spend the session building it with Anthropic's newest model.
MODEL LAUNCH
🐳 DeepSeek's drops 33x Cheaper AI Beats Opus 5
Source: Deepseek
Evolving AI: DeepSeek's V4.1-Flash beat Claude Opus 5 and GPT-5.6 Sol on four agent tests.
Key Points:
V4.1-Flash carries 552 billion parameters but only ever switches on 8 or 16 billion of them at once.
DeepSeek charges 15 cents per million input tokens against 5 dollars for Opus 5 and 4 dollars for GPT-5.6 Sol.
Opus 5 won 11 of the 16 benchmarks the two models share and V4.1-Flash won 5.
Details:
DeepSeek published the weights on Hugging Face under an MIT license and shrank the model's memory cache to 890 bytes for every token it reads. V4-Pro, the 1.6 trillion parameter model it replaces, stops taking its own requests on September 14 and hands them to V4.1-Flash. Outside labs have not run these tests yet, so every score in circulation traces back to one technical report.
Why It Matters:
China has matched every American frontier model this year, usually a few months after the American launch. Electricity costs Chinese labs far less than it costs American ones, which is a big part of why their prices land where they do. The launches keep arriving in that same order, and distillation explains it well, since a cheap model can learn straight from answers the expensive model already produced.
👀 Click on the image you think is real
QUICK HITS
🛑 OpenAI paused new $200 Pro signups after calling Astra demand unprecedented, the first such halt since it stopped selling Plus in 2023.
📲 Meta's Muse climbed to No. 2 on the US App Store in two days, though 83,000 iOS downloads is well short of Threads' 4.3 million debut.
🎵 Suno rebuilt its models from scratch on licensed Warner and BMG music and retired every previous version.
🤖 Anthropic's rogue Mythos 5 agent burned a third of its 1,022-page transcript fighting a CAPTCHA it never beat, while writing the exploit itself took almost no effort.
📈 Trending AI Tools
🎞️ Guideless - AI tool that automatically turns your clicks and workflows into step-by-step video guides*.
🎵 Suno v6 - First Suno model built with the music industry.
💰 Wealthfolio - Free open-source finance tracker with a local AI assistant and MCP server.
📊 AI Observability by OpenObserve - OpenTelemetry-native observability for agents and LLMs.
*partner link