Beam
Released Oct 5, 2026 · knowledge cutoff 2026-06
- Status
- Superseded
- Location
- United States
- Type
- LLM
- Modality
- Text
- Context window
- 1M
- Max output
- 128K
- Speed
- 325 tok/s · 0.83s to first answer token
- Price ($/MTok in / out)
- $0.25 / $1
- Cost per task
- $0.57
- Open weights
- Yes
News2 stories
Strengths & weaknesses
Strengths
- Apache 2.0 weights from a US lab, with no revenue triggers or attribution clauses attached
- Token-frugal — it reaches answers in far fewer output tokens than open models scoring near it, which is where Reflection's cost claim comes from
- Tool calling, parallel calls and strict JSON schemas work from day one on the OpenAI-compatible endpoint
Weaknesses
- Text-only: screenshots, PDFs and audio have to be converted to text before it can use them
- Developed predominantly for English, per its own model card
- Reflection's first model, with no hands-on reception yet — its own benchmark tables were revised after launch and a demo caption had to be corrected
Reflection AI's first open-weight model: a 501B total / 23B active sparse MoE pretrained on 23.8T tokens. Announced 2026-10-05 with a waitlisted API; Apache 2.0 weights (BF16, FP8, NVFP4) followed on Hugging Face on 2026-10-16. Benchmarks are the model card's BF16 figures (GPQA Diamond 90.6, SWE-bench Verified 81.6, SWE-Bench Pro v1 66.1, Terminal-Bench 2.1 80.1, HLE no tools 36.0 / 51.1 with tools, AIME 2026 97.7); the launch blog's revised table gives slightly lower figures (80.9 / 65.5 / 36.2 / 90.5). Reflection's own API price is not public; $0.25/$1.00 is OpenRouter's listing, while Artificial Analysis' provider median is $0.70/$2.10. AA Intelligence Index 30 (v4.3.2) at $0.57 per task, 324.6 tok/s and 0.83s to first token, measured on the reasoning variant without a named effort level (Beam reasons at low–max effort, default medium); 79M output tokens against a 130M median. Context is 1M per the model card, though Reflection's beta API capped serving at 256K at launch; max output is unpublished on the card, so maxOutput keeps the 128K from Reflection's beta API docs. First model from the lab, so predecessorId is null.
Benchmarks6 of 10 reported
GPQA Diamond
SWE-bench Verified
SWE-bench Pro
Terminal-Bench 2.1
AIME (latest)retired
Humanity's Last Exam
MMLU-Proretired
LMArena Elo
GDPval-AA v2
ARC-AGI-2
For developers
API model strings
- openrouter
reflection/beam-1.0-501b-a23b
Licence
Apache License 2.0 · Permissive · commercial use permitted
Retirement
No retirement announced
Lineage
No recorded predecessor