Reflection's Beam: strong on workflows, weak in the terminal

Beam, the first model from Reflection, is an open-weights mixture-of-experts model: it has 501 billion parameters in total, but only 23 billion are used for each token it processes. Artificial Analysis lists it as text-only, with a 1M-token context window and an Apache 2.0 license (Artificial Analysis model page). In that firm's independent testing Beam scores 30 on the Artificial Analysis Intelligence Index, which the evaluator says makes it the second-best US open-weights model, behind Meta's Muse Spark Open (39) and ahead of NVIDIA's Nemotron 3.5 Super VL (26). Outside China, only Muse Spark Open and the UAE's K2 Horizon (31) score higher (Artificial Analysis on X). Reflection's own launch post is a separate Intel item. Everything below comes from Artificial Analysis's measurements, not from Reflection's claims.

That single score hides an uneven profile. The index is a weighted average of ten evaluations: agent tasks count for 30%, coding for 20%, scientific reasoning for 20% and a general category for 30% (methodology). Beam does well on AutomationBench-AA, which is Artificial Analysis's run of Zapier's AutomationBench. In it, a model completes business workflows across simulated apps such as Gmail, Slack and Salesforce by finding and calling their REST APIs, and a task scores zero if the model breaks any guardrail. Beam reaches 48.5% there, 11 points ahead of K2 Horizon (37.2%) and level with Qwen3.8 27B (48.2%). Terminal-Bench 4.0 is the opposite case. It has 66 command-line tasks, run in the mini-swe-agent harness, and a task passes only if every one of its tests passes. Beam scores 3.0%, behind Qwen3.8 27B (5.6%) and Muse Spark Open (8.1%). Beam also scores 28 on Artificial Analysis's Coding Agent Index when run in Mirror, Reflection's own open-source command-line harness, 15 points behind the next setup.

Knowledge shows a similar trade-off. AA-Omniscience gives points for correct answers and subtracts points for wrong ones, while declining to answer costs nothing. Beam answers 33% of its questions correctly, well ahead of K2 Horizon (18%) and Qwen3.8 27B (16%). Its hallucination rate is 47%, though, against 26% and 30% for those two. Beam still ends with the best net score of the three, 1.5 against −3.0 and −10.0 (thread). In other words, it knows more than these peers, but it often guesses where it could have declined.

Beam's clearest strength is efficiency. It uses about 26,000 output tokens per index task, roughly 19,000 for reasoning and 7,000 for answers, while K2 Horizon uses about twice as many to score one point higher. Proprietary models remain well ahead: OpenAI's GPT-6 Astra (max) scores 53 on a similar budget of about 27,000 tokens. Per token, Beam is not cheap. The model page lists median provider prices of $0.70 per million input tokens and $2.10 per million output tokens, which Artificial Analysis calls somewhat expensive next to similar-sized open models (medians $0.43 and $1.68). It works out to $0.57 per index task (model page). All of these figures come from a single evaluator's harnesses, so they show how Beam performed under Artificial Analysis's setup, not every way the model can be run.