Reflection releases Beam's 501B open weights, scoring 30 on AA's index

Reflection released Beam, a 501B-parameter open-weight model that scores 30 on Artificial Analysis's index, making it the second-best US open-weight model behind Muse Spark Open's 39, according to Reflection. Beam uses about half the output tokens of K2 Horizon for a similar score, but carries a 47% hallucination rate and scores 3% on Terminal-Bench 4.0.

Beam is now the second-best US open-weight model, behind Muse Spark Open's 39, and uses about half the output tokens of K2 Horizon for a similar score. The cost is a 47% hallucination rate and 3% on Terminal-Bench 4.0. Read: Beam is now the second-best US open-weight model, behind Muse Spark Open's 39, and uses about half the output tokens of K2 Horizon for a similar score. The cost is a 47% hallucination rate and 3% on Terminal-Bench 4.0. Read: The NSA supplied these seeds in 2000 with no explanation, which fed years of backdoor suspicion around P-256 and its siblings. A clock-seeded generator anyone can rerun leaves little room for a hidden weakness, and a Claude model found it without hints in under 30 minutes. Read: Agent hours now outnumber employee hours 6 to 1 at Anthropic, and output per person rose as staff adopted agents. Reading all their agents' output would take non-technical staff half a work week, so human attention, not code generation, now limits what agents deliver. Read: At an equal token budget, a simple whole-file selector beat the best of three prompt compressors by 6.3 points and the best retriever by up to 12.7. Compressors alter the code text, so fewer patches apply, and retrievers often miss the files that must change. Read: Each connector used to cost engineers anywhere from half an hour to days of sign-ups, test accounts and waiting. An agent now moves every MCP server through vetting, OAuth discovery and two environment checks, while people still review each code change and handle partners. Read: Verified status used to rest on identity checks at approval. Cloudflare now measures whether a bot honors site preferences and responds to direct feedback, and will publish the scores on Radar, so crawlers and agents that ignore site requests lose trust in public.