ReflectionReleased Oct 5, 2026
Open weights

Beam

Strongest in Academic Reasoning (#58 of 228) and Knowledge Work (#62 of 186); weakest in Abstract Reasoning (#152 of 226). All seven dimensions are measured.

Composite IQ ?
114
#68 of 186 ranked models · ahead of 64%
US (Other)'s #2 model, behind Inkling 2 Small (125)
Effective cost
Not priced
No standard API rate
Speed
384tok/s
7s typical response
Context
1Mtokens
1,048,576 tokens in
Evidence
7of 7 dims
16 of 39 benchmarks

IQ profile

Seven dimensions, averaged equally into Composite IQ
Compare shape with
100 avg human130160● Abstract86#152● Math120#69● Academic128#58● Coding115#81● Computer Use113#81● Reliability116#73● Knowledge Work118#62
Measured95% range
#1 in dimensionTop-10 median

IQ dimensions

Ranks are among models scored in each dimension. Ranges are model-based estimates, not calibrated confidence intervals.

DimensionIQPosition · 70–160RankEvidencevs Top 10
86measured
#152 / 226
2/5
-46.5
120measured
#69 / 186
1/5
-29.5
128measured
#58 / 228
3/6
-7
115measured
#81 / 187
2/6
-36
113measured
#81 / 186
2/5
-33.5
116measured
#73 / 188
2/6
-19.5
118measured
#62 / 186
4/6
-14
Beam95% range (model-based, uncalibrated)Top-10 median#1 in dimension100 = average human

Benchmark results

16 of 39 Composite IQ benchmarks have a published result for Beam. Benchmarks without a result stay listed so gaps are visible.

Abstract Reasoning IQ 86 · measuredBest: #50 of 74 on ARC-AGI-22/5
BenchmarkScorePercentileRankLeader · source
ARC-AGI-2
Full source leaderboard reconciliation · Oct 23
23.5%
#50 / 74
Leader gpt-6.2-astra · 95.8%
Full source leaderboard reconciliation · Oct 23
ARC-AGI-1
Full source leaderboard reconciliation · Oct 23
74.0%
#51 / 74
Leader fable-5 · 98.5%
Full source leaderboard reconciliation · Oct 23
Not yet reported
ARC-AGI-3Chess PuzzlesMystery Game Puzzles
Mathematical Reasoning IQ 120 · measuredBest: #41 of 49 on ProofBench 1.11/5
BenchmarkScorePercentileRankLeader · source
ProofBench 1.1
Full source leaderboard reconciliation v1.1 · Oct 27
18.0%
#41 / 49
Leader fable-5.1 · 100%
Full source leaderboard reconciliation v1.1 · Oct 27
Not yet reported
FrontierMath Tier 4FrontierMath Tier 1-3MathArenaAIME
Academic Reasoning IQ 128 · measuredBest: #49 of 159 on CritPt3/6
BenchmarkScorePercentileRankLeader · source
CritPt
Artificial Analysis v4.3.2 · Oct 27
13.1%
#49 / 159
Leader Muse Wonder Full · 32.9%
Artificial Analysis v4.3.2 · Oct 27
Humanity's Last Exam
Artificial Analysis v4.3.2 · Oct 27
34.7%
#70 / 167
Leader opus-5.5 · 61.4%
Artificial Analysis v4.3.2 · Oct 27
SciCode
Artificial Analysis v4.3.2 · Oct 27
48.5%
#72 / 167
Leader opus-5.5 · 66.9%
Artificial Analysis v4.3.2 · Oct 27
Not yet reported
GPQA DiamondMMLU-ProMMMU-Pro
Programmatic Reasoning IQ 115 · measuredBest: #54 of 109 on Terminal-Bench 4.02/6
BenchmarkScorePercentileRankLeader · source
Terminal-Bench 4.0
Artificial Analysis v5.0.1 · Nov 13
3.0%
#54 / 109
Leader Kimi K3.1 Pro · 67.2%
Artificial Analysis v5.0.1 · Nov 13
IOI V2
Full source leaderboard reconciliation v2 · Oct 27
22.8%
#39 / 43
Leader Gemini 4 Argon · 100%
Full source leaderboard reconciliation v2 · Oct 27
Not yet reported
LiveCodeBenchProgramBenchSWE-rebenchMirrorCode
Computer Use IQ 113 · measuredBest: #25 of 45 on BrowseComp2/5
BenchmarkScorePercentileRankLeader · source
BrowseComp
Full source leaderboard reconciliation · Oct 23
76.8%
#25 / 45
Leader Kimi K3.1 Pro · 92.1%
Full source leaderboard reconciliation · Oct 23
Arena.ai Agent Arena
Arena.ai Agent Arena · Nov 19
-8.3
#63 / 77
Leader fable-5.2 · 27.4
Arena.ai Agent Arena · Nov 19
Not yet reported
OSWorld-VerifiedMCP AtlasAgents' Last Exam
Reliability IQ 116 · measuredBest: #63 of 158 on AA Omniscience2/6
BenchmarkScorePercentileRankLeader · source
AA Omniscience
Artificial Analysis v4.3.2 · Oct 27
1.5
#63 / 158
Leader opus-5.5 · 46.4
Artificial Analysis v4.3.2 · Oct 27
AA Long Context Reasoning v1.1
Artificial Analysis v4.3.2 · Oct 27
78.0%
#68 / 154
Leader kimi-k3 · 88.7%
Artificial Analysis v4.3.2 · Oct 27
Not yet reported
SimpleQA VerifiedBullshitBench v2IFBenchFACTS Grounding
Knowledge Work IQ 118 · measuredBest: #44 of 112 on AutomationBench-AA4/6
BenchmarkScorePercentileRankLeader · source
AutomationBench-AA
Artificial Analysis v5.0.1 · Nov 13
48.5%
#44 / 112
Leader Gemini 4 Carbon · 77.8%
Artificial Analysis v5.0.1 · Nov 13
GDPval-AA v2.1 normalized score
Artificial Analysis v4.3.2 · Oct 23
33.9%
#54 / 124
Leader opus-5.5 · 68.3%
Artificial Analysis v4.3.2 · Oct 23
AA-Briefcase v1.1 Elo
Artificial Analysis v5.0.1 · Nov 13
976
#58 / 116
Leader sonnet-5.5 · 1,816
Artificial Analysis v5.0.1 · Nov 13
GDP.pdf
Artificial Analysis v5.0.1 · Nov 13
11.6%
#57 / 111
Leader Muse Wonder Full · 32.6%
Artificial Analysis v5.0.1 · Nov 13
Not yet reported
AA-AnalystAgentEnterpriseOps-Gym-AA

Release & availability

Facts from the publisher's announcement and documentation.

Released
Developer
Reflection
Access
Apache 2.0 open weights are now available in BF16, FP8 and NVFP4; hosted access through Reflection and inference partners.
Weights
Open weightsDownloadable and self-hostable
Inputs
text
Context
1,048,576 tokens
API price
No per-token rateCurrent standard hosted API pricing is not published in the checked official sources.

What Reflection says

  • Reflection announced Beam on October 5, 2026 and released the weights on October 16.
  • Text-only mixture-of-experts model with 501B total and 23B active parameters, supporting a 1M-token context.

Limitations

  • Predominantly developed for English.
  • Publisher benchmark tables are self-reported; only exact compatible independent source rows are imported.
  • Standard hosted token pricing remains unverified.

Publisher statements, not AI IQ measurements.

About this score

How much of Beam's IQ is measured, and how the rest is estimated.

16/ 39benchmark results
7/ 7dimensions measured
0estimated
2/5Abstract
1/5Math
3/6Academic
2/6Coding
2/5Computer Use
2/6Reliability
4/6Knowledge Work
Bar height = share of benchmarks with resultsEstimatedNo score
Measured results are shown as published

Every score in the benchmark tables comes from its source unchanged. Composite IQ averages all seven dimensions equally.

How gaps are handled

Every dimension has at least one published result.

A missing result is not a zero

Benchmark versions and evaluation harnesses can differ between sources, so small gaps between models may not be meaningful.

Read the scoring methodology →

Compare Beam

Side-by-side pages for the models it is most often weighed against.

BeamComparison model
Browse all comparisons →

Frequently asked questions

Quick answers about Beam, drawn from the same data as the rest of this page.

How AI IQ scores models →Get this data from the API →
What should I know about Beam?
Originally announced October 5, 2026; public weight availability verified October 16 under Apache 2.0. Current API ID Beam-1.0-501B-A23B supports a 1,048,576-token combined context and mandatory reasoning; publisher defaults may differ by organization. BF16, FP8 and NVFP4 are numerical deployment formats, not additional canonical identities. Standard hosted prices and raw speed are unverified; no local/free inference price is substituted.
What is the IQ of Beam?
Beam has an estimated IQ of 114, which ranks #68 out of 186 public models on the AI IQ leaderboard. The score combines seven equally weighted reasoning dimensions using measured benchmark results and conservative estimates for missing coverage.
What is Beam best at?
Among the dimensions with published benchmark results, Beam ranks highest in Academic Reasoning (IQ 128, #58 of 228) and Knowledge Work (IQ 118, #62 of 186). Its lowest-ranked measured dimension is Abstract Reasoning (IQ 86, #152 of 226).
What is the EQ of Beam?
Beam doesn't have an EQ score yet. EQ (emotional reasoning) is calculated from EQ-Bench 4 and AttuneBench, and neither has published a result for Beam. AI IQ doesn't estimate EQ from other benchmarks, so the score appears once one of those results is published.
Which benchmarks has Beam been tested on?
Beam has published results on 16 of the 39 benchmarks behind AI IQ's Composite IQ, including CritPt, AutomationBench-AA, AA Omniscience, Humanity's Last Exam and SciCode.
How much does Beam cost?
AI IQ doesn't have a verified standard per-token API price for Beam, so it can't calculate an effective cost. This is common for models offered only through subscriptions, restricted-access programs, or self-hosting.
Is Beam open source?
Yes — Beam is an open-weights model: its trained weights are published, so it can be downloaded and self-hosted.
Who developed Beam?
Beam was developed by Reflection, based in the United States.
When was Beam released?
Beam was publicly released on October 5, 2026.
How large is the context window of Beam?
Beam supports a context window of 1,048,576 tokens — the amount of text it can consider in a single request.
How fast is Beam?
In independent measurements, Beam generates about 384 tokens per second, with a typical response time of about 7.1 seconds.