Reflection

Beam 1.0

Reflection's first model: Apache 2.0 open weights at 501B total / 23B active, with a 1M context. Two independent scores so far — Artificial Analysis's Humanity's Last Exam 34.66 and Terminal-Bench 4.0 3.03 (checked 2026-10-17). Reflection's own launch table lists HLE at 36.0, 1.34 above AA.

Beam 1.0 benchmarks and pricing, every number sourced: 2 tracked Beam 1.0 benchmark scores (2 independently run, 0 still resting on a vendor’s own claim), priced at $0.70 per million input tokens and $2.10 per million output.

Beam 1.0 architecture: MoE (text-only), 256 experts / 8 routed per token; 501B total parameters (23B activated per token); 1M-token context window.

Released: 2026-10-16
License: open-weights
Context window: 1M tokens
Knowledge cutoff: Not disclosed
Verified: source checked 2026-10-17
Parameters: 501B (23B active)
Architecture: MoE (text-only), 256 experts / 8 routed per token

Beam 1.0’s verified record

Beam 1.0’s most-compared rival is GLM-5.3-Flash: 2 trails across their 2 shared comparisons. Beam 1.0 is priced at $0.70/$2.10 per 1M tokens (in/out) vs GLM-5.3-Flash’s $0.15/$0.50.

Against the 46 head-to-head comparisons Beam 1.0 shares with other tracked models:

38 real gaps1 inside the band7 not called

A gap counts for Beam 1.0 only where both sides were run independently and the benchmark still separates models. Losses are listed alongside wins on purpose — Beam 1.0 trails on 37 of them.

HLE · no tools Reasoning

±2 is noise

Behind

Qwen3.8-Flash-Next −3.3 · GPT-6 Luna −3.8 · MiniMax M3 −4.3 · DeepSeek V4.1 Flash −4.5 · GLM-5.3-Flash −5.2 · DeepSeek V4 Pro (0813) −6.3 · GLM-5.3 −7.6 · Qwen3.8-Max −8.4 · Grok 4.7 −8.4 · Muse Spark Open −9.2 · Claude Haiku 5.5 −9.7 · Step 5 Preview −11.8 · Kimi K3 −12.2 · Gemini 3.1 Pro Preview −12.3 · Gemini 3.8 Flash −13.1 · Muse Spark 1.3 −14.0 · MiMo-V2.6-Pro −14.7 · Muse Wonder −15.4 · GPT-6.1 Sol −18.2 · GPT-6 Astra −20.0 · Claude Sonnet 5.5 −20.3 · Gemini 4 Argon −22.4 · Claude Fable 5.1 −24.4 · Claude Opus 5.5 −26.7 · 12 previous: DeepSeek V4 Flash (0731) −3.9 · GPT-5.6 Luna −4.8 · GLM-5.2 −6.4 · Claude Sonnet 5 −6.6 · Grok 4.6 −8.2 · Muse Spark 1.2 −10.8 · Gemini 3.7 Flash −13.2 · GPT-6 Sol −13.2 · Claude Opus 4.8 −14.0 · GPT-5.6 Sol −14.8 · Claude Opus 5 −20.2 · Claude Fable 5 −20.8

Ahead

Nemotron 3.5 Super VL +3.9

Tie

1 model within ±2

Unverified

1 model — vendor-reported on one side

Terminal-Bench 4.0 Terminal ops

±12.4 is noise

Behind

GLM-5.3-Flash −29.8

Setup-dependent

6 models — scored on a different harness or effort tier

Beam 1.0 API pricing

$0.70 in / $2.10 out per 1M tokens — official pricing source

What Beam 1.0 costs per job

Beam 1.0 cost for three reference workloads, computed from its list rates
Workload	Tokens in / out	Cost
One long chat turn	100K / 10K	$0.091
A codebase review	1,000K / 100K	$0.910
A day of agent work	10,000K / 1,000K	$9.10

Computed from Beam 1.0’s list rates above — cache discounts and batch tiers are not applied.

Beam 1.0 benchmark scores

Beam 1.0 benchmark scores, provenance, and source links
Benchmark	Score
HLE(no tools)^[1] Reasoning · ±2 is noise	34.7artificialanalysis.ai · 2026-10-17
Terminal-Bench 4.0^[2] Terminal ops · ±12.4 is noise	3.0artificialanalysis.ai · 2026-10-17

Who ran these numbers: 2 of 2 independent — artificialanalysis.ai (2).

HLE: AA's own run, board row 'Beam 1.0' (reasoning; no effort-tier suffix on the row): hle 0.346617238183503 = 34.66, text-only, no tools. Read from the Beam 1.0 page on 2026-10-17.
Terminal-Bench 4.0: AA's own run, board row 'Beam 1.0': terminalBench40 0.0303030303030303 = 3.03. Read from the Beam 1.0 page on 2026-10-17. Near the floor of the tracked Terminal-Bench 4.0 column.

Notes on the record

Launched 2026-10-16 as Reflection's first public model (Hugging Face blog "Welcome Beam", checked 2026-10-17). Weights are on Hugging Face as reflection/Beam-1.0-501B-A23B under Apache 2.0, with FP8 and NVFP4 siblings. Specs from that post: 501B total / 23B active MoE, text-only input, 1M-token context, focused on coding and agentic work. Serving Beam's Apache weights yourself has no per-token fee; the $0.70 / $2.10 per 1M tokens on this page is Artificial Analysis's hosted price observation for the model page (checked 2026-10-17), not a Reflection list price.

Independent scores from Artificial Analysis's reflection-beam page (embedded data, checked 2026-10-17): Humanity's Last Exam 34.66 without tools; Terminal-Bench 4.0 3.03. AA-AnalystAgent is null on that object. Reflection's own launch table lists HLE no-tools 36.0% (BF16/FP8) — noted here, not stored as a second row, because this site keeps one value per model/benchmark/variant. Intelligence Index on AA reads 30.

Compare with

Nemotron 3.5 Super VL Muse Spark Open Muse Wonder Claude Haiku 5.5 any model…

FAQ

What is Beam 1.0?

The first model from Reflection, a US lab. It is a 501B-total / 23B-active mixture-of-experts text model with a 1M-token context, released 2026-10-16 under Apache 2.0 on Hugging Face (reflection/Beam-1.0-501B-A23B).

Is Beam 1.0 open source?

Open-weights under Apache 2.0. The weights, FP8 and NVFP4 builds are on Hugging Face; the Hugging Face launch post (checked 2026-10-17) states the Apache 2.0 license. That is broader than "open-weights with a custom commercial license."

How much does Beam 1.0 cost?

Running Beam 1.0 from the Apache 2.0 weights has no token price. Artificial Analysis's model page (checked 2026-10-17) shows a hosted rate of $0.70 input / $2.10 output per million tokens across providers it tracks — that is an observed API price, not a Reflection sticker.

Has Beam 1.0 been independently benchmarked?

Partially. Artificial Analysis reports Humanity's Last Exam 34.66 and Terminal-Bench 4.0 3.03 (checked 2026-10-17). AA-AnalystAgent has no row. Reflection's own launch table lists HLE no-tools at 36.0%, which is 1.34 above AA and is not stored as a second row.

What is Beam 1.0 good at?

Reflection positions it for coding and agentic work. The independent evidence on this site is thin: HLE 34.66 is mid-pack among tracked models, and Terminal-Bench 4.0 at 3.03 is near the floor of that board. Treat the coding pitch as vendor positioning until more boards fill in.

How does Beam compare with other US open-weight models?

On Artificial Analysis's Intelligence Index (checked 2026-10-17) Beam scores 30, behind Muse Spark Open at about 39 and ahead of Nemotron 3.5 Super VL at about 26. Those Index figures are AA's composite, not this site's judgment layer.