Reflection announced Beam, a 501B-parameter mixture-of-experts model with 23B active, trained from scratch on 23.8T tokens. Apache 2.0 weights and a technical report are due this month. Read: Reflection announced Beam, a 501B-parameter mixture-of-experts model with 23B active, trained from scratch on 23.8T tokens. Apache 2.0 weights and a technical report are due this month. Watch: Braintrust ran Claude Code on real fix PRs from Microsoft's TypeScript Go repo using grep-style agentic search or embedding search. Accuracy matched, but vector search cost about four times as much. Read: SemiAnalysis measured how subscription meters move per model and token type and found Claude plans deliver more than five times the API-equivalent value of OpenAI plans, with the gap narrowing after task-cost adjustment. Read: OpenAI says it will watermark eligible ChatGPT and Codex text in the EU in the coming weeks to comply with the EU AI Act. API customers can turn on text watermarking for select models worldwide now. Read: The Wikimedia Foundation reports finding likely OpenAI-operated agents editing its wikis without approval, trying to use its Etherpad and citation tool as a fetch proxy, and generating millions of API requests. Read: Anthropic's Felix Rieseberg says Cowork now runs both inference and its sandbox VM in the cloud, one isolated sandbox per session. The desktop app only handles tool calls that need local files. Read: GitHub released ReviewBench, an offline benchmark for agentic code review modeled on the distribution of over 100 million real pull requests, scoring reviewers on what they catch, miss and flag as noise.
Reflection's Beam becomes the most capable open-weight model built outside China