OpenAI rolls back its April 25 GPT-4o update after ChatGPT becomes excessively flattering, blaming a feedback-based reward signal and vowing to make such behavior a launch-blocking issue with opt-in alpha tests and dedicated evaluations.
Basecamp Research raises a $140 million Series C to train AI models for designing therapeutics, beginning with in vivo cell therapy, in a round led by S32 with Nvidia’s NVentures and Anthropic’s Anthology Fund participating.
Anthropic cuts Claude Opus 5.5 token prices—20% for input and output and 60% for cache reads—saying a representative Claude Code session now costs $2.40 versus $3.50 on Opus 5.
OpenAI launches GPT-6 Sol and Luna at roughly half the prior generation’s price, with Sol costing $2 per million input tokens and $10 per million output tokens while claiming benchmark wins over Claude Opus 5.
Anthropic unveils Claude Opus 5.5 as its strongest behavioral-alignment model, restricting its use for cybersecurity and biology tasks to vetted organizations while pricing it at $4 per million input tokens and $20 per million output tokens.
Alibaba unveils its T-Head-designed Zhenwu V900 AI accelerator, claiming it is China’s most powerful chip and can be clustered in systems of up to 500,000 units to train frontier models in a challenge to Nvidia.
Anthropic and OpenAI slash flagship AI model prices—Claude Opus 5.5 costs 40% less and GPT-6 Sol and Luna half as much—while OpenAI says an internal system has solved more than 100 open math problems since Aug. 28.
OpenAI says its internal model solves more than 100 additional open mathematics problems and creates an independent IAS-based advisory group of nine mathematicians, who will assess results and can publicly disagree with the company.
Xiaomi unveils MiMo-V2.6, an omni-modal AI model family it says significantly improves coding, design and cybersecurity through groupwise reinforcement learning, with its Pro version reaching 72.6 average@3 on DeepSWE v1.1 after the team freezes an unstable Mixture-of-Experts router.
Bespoke Labs launches Nimble, an open 9B-parameter local model that makes typed boolean or enum text decisions with answer probabilities, matching 292 of 324 held-out labels while limiting inputs to 2,048 tokens and schemas to 26 choices.