This report relies on a secondary write-up of Reflection's announcement and API documentation (eesel.ai, dated 8 October 2026). We could not open Reflection's own pages. Treat all figures as vendor-reported.
Training scale, as claimed: pretraining on 23.8T tokens using 6,144 Nvidia GB300 GPUs in under four weeks, followed by reinforcement learning on 10.5K GB300 GPUs for four weeks with more than 100 million rollouts.
Context is described two ways. Training was said to reach 1M effective tokens. The beta API serves 256K input and 128K maximum output, and that may change. Input is text only; the API does not accept images, audio or files.
Reflection published a benchmark table comparing Beam with GLM-5.2. Beam scores: SWE-bench Pro 65.5 against 62.1, Terminal Bench v2.1 80.1 against 81.0, DeepSWE 44.4 against 44.0, MCP Atlas 78.7 against 77.8, tau3 banking 38.0 against 37.1, HLE without tools 36.2 against 40.5, and GPQA Diamond 90.5 against 91.2. It also lists SWE-bench Verified at 80.9 with no GLM-5.2 figure. The differences are small in both directions.
On efficiency, Reflection says Beam matches GLM-5.2 on reasoning with 3-4x less inference compute. Its own caveat says these are approximate compute estimates and not measured cost.
Access today is a waitlisted API with an OpenAI-compatible endpoint. A pricing page exists but sits behind login, so the price is unknown. Reflection has not published weights. A promised release includes FP8 and NVFP4 versions, a technical report, a model card and developer tooling.
Why this is not in the model registry: the NextVector model list only accepts entries with a source key such as an Arena row, an OpenRouter id or an official price page. Beam has none yet. We will add it when the weights and an independent score exist.