Anthropic says it is very close to recursive self-improvement capability

Anthropic says AI is nearing the point where it trains itself well enough for capabilities to accelerate dramatically. Alibaba's Qwen head commits to open-weight Qwen 4.0 and 5.0 models. Unconfirmed reports point to new OpenAI and Qwen models on the way.

1. Anthropic warns it is very close to recursive self-improvement capability

   A frontier lab saying AI may soon train itself well enough to speed up capability gains dramatically is a major signal about where development is heading.

2. Qwen head says future Qwen 4.0 and 5.0 models will stay open source

   Liu Dayi rejected rumors that Qwen4 would be the last open series; Qwen already counts over 3 billion downloads and more than 300,000 derivatives.

3. Astra can read and write model weights to improve Qwen

4. ISTA releases expert-pruned DeepSeek-V4.1-Flash-1017 GSQ-RCO Coder at about 1 bit per parameter

   Pruning half the routed experts cuts weights to 72.1 GB from 138.2 GB while keeping 94.8% of SWE-bench Verified performance, and it runs in llama.cpp.

5. GPT-6.2 Astra spotted in Codex usage analytics ahead of possible launch

   Reported but unconfirmed: a GPT-6.2-astra model appearing for some Codex users suggests OpenAI may already be serving it quietly.

6. Alibaba Qwen Code repo reveals qwen4-max, qwen4-coder-plus and qwen4-omni-120b model IDs

   Unannounced IDs in Qwen Code pull requests, reportedly tied to a 262,144 context limit, hint at new Qwen4 models; this remains a rumor.

7. Anthropic adds Claude Code rule against recreating existing video games

   Claude Code 2.1.314 now tells Claude not to rebuild specific existing games even when renamed, and XDA found Claude Opus 5.5 has separate limits too.

   Models: Claude Opus 5.5

8. LongCat-2.5 from Meituan's LongCat team is a natively multimodal model for agentic tasks

   Meituan's LongCat team is pointing a natively multimodal model at agentic workloads, widening the field of companies building for agents.

Written by Claude (claude-opus-5-5) from the stories below.