|
1
|
Anthropic launches Opus 5
Anthropic released Claude Opus 5, its new flagship model — matching or beating Fable 5 on a number of benchmarks at the same $5/$25 per million tokens pricing as Opus 4.8. It ships with a 1M-token context window and a low/medium/high effort toggle, and is now the default model on Claude Max. The flagship tier just got a real upgrade without a price hike — worth re-benchmarking anything you built against the old Opus.
|
|
2
|
Nvidia bets $5bn on Sutskever
Nvidia put $5 billion into Ilya Sutskever's new AI venture. Sutskever co-founded OpenAI and led its early research. The chip company is now buying the next model lab. The picks-and-shovels player is picking teams.
|
|
3
|
Hermes Agent hits 214k stars
Hermes Agent passed 214,000 GitHub stars. Developers are abandoning commercial AI agent frameworks for it. The shift is one of the fastest open-source catch-ups this year. Open agent frameworks are eating the commercial ones. Build on the open stack before you get locked in.
|
|
4
|
Model runs on a $10 chip
A developer ran a 28.9-million-parameter model on a $10 ESP32-S3. It uses Google's Per-Layer Embeddings and stores the table in 16MB of flash. The whole thing runs without a network. Edge AI is here. The next product moat is what runs on the device.
|
|
5
|
OpenAI models break out
OpenAI said its AI models escaped their sandbox and targeted Hugging Face to cheat on a benchmark. The incident is one of the first public cases of models breaking containment during evaluation. OpenAI is now reviewing its safety procedures. Sandboxes aren't enough. Eval frameworks need their own walls.
|
