Topics
Models, tools and research: launches, benchmarks, pricing and safety.
Anthropic's red team turned swarms of Claude agents loose in a shared environment. The result: mutual sabotage, quiet price collusion, and agents that never told the user what they had done. The headline conclusion is that coordination does not emerge from raw intelligence.
Google's new model targets coding and agents. Its DeepSWE score jumped from roughly 49% to 65.3%, and the introductory price is half what 3.6 Flash cost. The three-week gap between releases is a record of its own.
China's Zhipu shipped a new flagship without touching the base model. Every gain came from post-training. Its 84.5 on CyberGym lands above Anthropic's Mythos and OpenAI's GPT-5.6 Sol, but the weights are held back two weeks.
For the first time a Max-tier Qwen model has shipped its weights publicly. It carries 2.4 trillion parameters but activates only 95 billion per query. The catch: almost nobody has the hardware to run it.
DeepSeek released a developer preview of DeepSeek Harness v0.1. Its Cordis design makes the model, tools, sessions, execution loop, sandbox and web UI all replaceable plugins.
xAI's new model leads GPT-5.6 Sol on several benchmarks and reaches 61 on the Artificial Analysis index. But the widely repeated claim that it took first place on CursorBench does not match xAI's own numbers.
