Claude Will Build Its Own Successor

Anthropic projects Claude will lead 80-90% of frontier model training by next year, up from just 3% in April.

OpenAI upgraded GPT-6 Astra Pro so multiple agents can split a task and work in parallel inside ChatGPT.

Inside Anthropic, Claude writes about 90% of Claude Code's code, leaving engineers to handle the rest.

Merged code changes more than doubled from June to September while reverted fixes fell over 40%.

Anthropic now has agents reviewing agents' code, with one pull request going through 50+ overnight rounds before a human stepped in.

Four Google DeepMind Gemini 4 Argon agents beat Minecraft from scratch in 27 hours with zero human help, writing 2,783 programs.

Claude reportedly solved a 25 year old NSA cryptography puzzle worth a $36,864 bounty in under 30 minutes.

Perplexity open sourced two new models, pricing one at a single cent per million tokens while topping a public leaderboard.

A side project started by a Berkeley PhD student to speed up Meta's open model now runs on roughly 500,000 GPUs as vLLM.

Open weight models now handle 72% of all tokens processed on OpenRouter, though closed models still take most of the money.