ModelsBeam

Beam

Reflection AI’s text-only mixture-of-experts model for coding, reasoning and agentic work, available as an early-access preview.

Reflection AIReleased 5 October 2026Open weights pending
Context window
1M tokens (trained)
Maximum output
128k tokens (beta API)
Input / output
Not publishedUSD per million tokens
Tests with results
2121 source-specific results
AA Intelligence Index
No resultNot published

Test coverage

  • 21self
  • 144 no result

165 catalogue tests. Each test counts once; evaluator results take precedence where sources overlap.

Results

Filter the recorded results; open Details on a row for run conditions.

21 results

BenchmarkResultSourceRun
DeepSWE v1.144.4%self
Details

Reflection AI’s four capability tables, read 6 October 2026. All 21 rows and eight model columns are transcribed; NR means not reported and remains blank. The source’s Qwen 3.8 Max column uses this catalogue’s Qwen3.8 Max spelling. Beam results are publisher reports; Reflection AI cites Artificial Analysis and Datacurve for comparison-model evaluations. The announcement does not specify a common harness, effort or trial count for every row. Ranks are derived within each published row from reported values only, with ties sharing a rank; they do not establish an independent or overall model ranking. AutomationBench uses the public split; AA-Omniscience is an index on the public split only, calculated by Reflection AI. BrowseComp and DeepSearchQA use context management. AA-LCR has no stated version. These settings remain separate from other source fields.

Metric
DeepSWE v1.1
Published
5 October 2026
Effort
Not reported separately
Tasks / attempts
Unknown / Unknown
Confidence interval
Not reported
Build a reproduction plan →
SWE-Bench Pro V2 HARD77.2%self
Details

Reflection AI’s four capability tables, read 6 October 2026. All 21 rows and eight model columns are transcribed; NR means not reported and remains blank. The source’s Qwen 3.8 Max column uses this catalogue’s Qwen3.8 Max spelling. Beam results are publisher reports; Reflection AI cites Artificial Analysis and Datacurve for comparison-model evaluations. The announcement does not specify a common harness, effort or trial count for every row. Ranks are derived within each published row from reported values only, with ties sharing a rank; they do not establish an independent or overall model ranking. AutomationBench uses the public split; AA-Omniscience is an index on the public split only, calculated by Reflection AI. BrowseComp and DeepSearchQA use context management. AA-LCR has no stated version. These settings remain separate from other source fields.

Metric
SWE Bench Pro v2-Hard
Published
5 October 2026
Effort
Not reported separately
Tasks / attempts
Unknown / Unknown
Confidence interval
Not reported
Build a reproduction plan →
SWE-Bench Pro v165.5%self
Details

Reflection AI’s four capability tables, read 6 October 2026. All 21 rows and eight model columns are transcribed; NR means not reported and remains blank. The source’s Qwen 3.8 Max column uses this catalogue’s Qwen3.8 Max spelling. Beam results are publisher reports; Reflection AI cites Artificial Analysis and Datacurve for comparison-model evaluations. The announcement does not specify a common harness, effort or trial count for every row. Ranks are derived within each published row from reported values only, with ties sharing a rank; they do not establish an independent or overall model ranking. AutomationBench uses the public split; AA-Omniscience is an index on the public split only, calculated by Reflection AI. BrowseComp and DeepSearchQA use context management. AA-LCR has no stated version. These settings remain separate from other source fields.

Metric
SWE Bench Pro v1
Published
5 October 2026
Effort
Not reported separately
Tasks / attempts
Unknown / Unknown
Confidence interval
Not reported
Build a reproduction plan →
Terminal-Bench 2.180.1%Terminal Bench v2.1self
Details

Reflection AI’s four capability tables, read 6 October 2026. All 21 rows and eight model columns are transcribed; NR means not reported and remains blank. The source’s Qwen 3.8 Max column uses this catalogue’s Qwen3.8 Max spelling. Beam results are publisher reports; Reflection AI cites Artificial Analysis and Datacurve for comparison-model evaluations. The announcement does not specify a common harness, effort or trial count for every row. Ranks are derived within each published row from reported values only, with ties sharing a rank; they do not establish an independent or overall model ranking. AutomationBench uses the public split; AA-Omniscience is an index on the public split only, calculated by Reflection AI. BrowseComp and DeepSearchQA use context management. AA-LCR has no stated version. These settings remain separate from other source fields.

Metric
Terminal Bench v2.1
Published
5 October 2026
Effort
Not reported separately
Tasks / attempts
Unknown / Unknown
Confidence interval
Not reported
Build a reproduction plan →
SWE Atlas Codebase QnA34.6%self
Details

Reflection AI’s four capability tables, read 6 October 2026. All 21 rows and eight model columns are transcribed; NR means not reported and remains blank. The source’s Qwen 3.8 Max column uses this catalogue’s Qwen3.8 Max spelling. Beam results are publisher reports; Reflection AI cites Artificial Analysis and Datacurve for comparison-model evaluations. The announcement does not specify a common harness, effort or trial count for every row. Ranks are derived within each published row from reported values only, with ties sharing a rank; they do not establish an independent or overall model ranking. AutomationBench uses the public split; AA-Omniscience is an index on the public split only, calculated by Reflection AI. BrowseComp and DeepSearchQA use context management. AA-LCR has no stated version. These settings remain separate from other source fields.

Metric
SWE Atlas Codebase QnA
Published
5 October 2026
Effort
Not reported separately
Tasks / attempts
Unknown / Unknown
Confidence interval
Not reported
Build a reproduction plan →
SWE-bench Multilingual78.0%self
Details

Reflection AI’s four capability tables, read 6 October 2026. All 21 rows and eight model columns are transcribed; NR means not reported and remains blank. The source’s Qwen 3.8 Max column uses this catalogue’s Qwen3.8 Max spelling. Beam results are publisher reports; Reflection AI cites Artificial Analysis and Datacurve for comparison-model evaluations. The announcement does not specify a common harness, effort or trial count for every row. Ranks are derived within each published row from reported values only, with ties sharing a rank; they do not establish an independent or overall model ranking. AutomationBench uses the public split; AA-Omniscience is an index on the public split only, calculated by Reflection AI. BrowseComp and DeepSearchQA use context management. AA-LCR has no stated version. These settings remain separate from other source fields.

Metric
SWEBench Multilingual
Published
5 October 2026
Effort
Not reported separately
Tasks / attempts
Unknown / Unknown
Confidence interval
Not reported
Build a reproduction plan →
SWE-bench (Verified)80.9%self
Details

Reflection AI’s four capability tables, read 6 October 2026. All 21 rows and eight model columns are transcribed; NR means not reported and remains blank. The source’s Qwen 3.8 Max column uses this catalogue’s Qwen3.8 Max spelling. Beam results are publisher reports; Reflection AI cites Artificial Analysis and Datacurve for comparison-model evaluations. The announcement does not specify a common harness, effort or trial count for every row. Ranks are derived within each published row from reported values only, with ties sharing a rank; they do not establish an independent or overall model ranking. AutomationBench uses the public split; AA-Omniscience is an index on the public split only, calculated by Reflection AI. BrowseComp and DeepSearchQA use context management. AA-LCR has no stated version. These settings remain separate from other source fields.

Metric
SWEBench Verified
Published
5 October 2026
Effort
Not reported separately
Tasks / attempts
Unknown / Unknown
Confidence interval
Not reported
Build a reproduction plan →
AIME 202697.8%self
Details

Reflection AI’s four capability tables, read 6 October 2026. All 21 rows and eight model columns are transcribed; NR means not reported and remains blank. The source’s Qwen 3.8 Max column uses this catalogue’s Qwen3.8 Max spelling. Beam results are publisher reports; Reflection AI cites Artificial Analysis and Datacurve for comparison-model evaluations. The announcement does not specify a common harness, effort or trial count for every row. Ranks are derived within each published row from reported values only, with ties sharing a rank; they do not establish an independent or overall model ranking. AutomationBench uses the public split; AA-Omniscience is an index on the public split only, calculated by Reflection AI. BrowseComp and DeepSearchQA use context management. AA-LCR has no stated version. These settings remain separate from other source fields.

Metric
AIME 2026
Published
5 October 2026
Effort
Not reported separately
Tasks / attempts
Unknown / Unknown
Confidence interval
Not reported
Build a reproduction plan →
Humanity's Last Exam36.2%HLE no toolsself
Details

No tools; distinct from tool-enabled HLE publisher rows. Reflection AI’s four capability tables, read 6 October 2026. All 21 rows and eight model columns are transcribed; NR means not reported and remains blank. The source’s Qwen 3.8 Max column uses this catalogue’s Qwen3.8 Max spelling. Beam results are publisher reports; Reflection AI cites Artificial Analysis and Datacurve for comparison-model evaluations. The announcement does not specify a common harness, effort or trial count for every row. Ranks are derived within each published row from reported values only, with ties sharing a rank; they do not establish an independent or overall model ranking. AutomationBench uses the public split; AA-Omniscience is an index on the public split only, calculated by Reflection AI. BrowseComp and DeepSearchQA use context management. AA-LCR has no stated version. These settings remain separate from other source fields.

Metric
HLE no tools
Published
5 October 2026
Effort
Not reported separately
Tasks / attempts
Unknown / Unknown
Confidence interval
Not reported
Build a reproduction plan →
SciCode49.7%self
Details

Reflection AI’s four capability tables, read 6 October 2026. All 21 rows and eight model columns are transcribed; NR means not reported and remains blank. The source’s Qwen 3.8 Max column uses this catalogue’s Qwen3.8 Max spelling. Beam results are publisher reports; Reflection AI cites Artificial Analysis and Datacurve for comparison-model evaluations. The announcement does not specify a common harness, effort or trial count for every row. Ranks are derived within each published row from reported values only, with ties sharing a rank; they do not establish an independent or overall model ranking. AutomationBench uses the public split; AA-Omniscience is an index on the public split only, calculated by Reflection AI. BrowseComp and DeepSearchQA use context management. AA-LCR has no stated version. These settings remain separate from other source fields.

Metric
SciCode
Published
5 October 2026
Effort
Not reported separately
Tasks / attempts
Unknown / Unknown
Confidence interval
Not reported
Build a reproduction plan →
CritPt16.3%CriPT AAself
Details

Reflection AI’s four capability tables, read 6 October 2026. All 21 rows and eight model columns are transcribed; NR means not reported and remains blank. The source’s Qwen 3.8 Max column uses this catalogue’s Qwen3.8 Max spelling. Beam results are publisher reports; Reflection AI cites Artificial Analysis and Datacurve for comparison-model evaluations. The announcement does not specify a common harness, effort or trial count for every row. Ranks are derived within each published row from reported values only, with ties sharing a rank; they do not establish an independent or overall model ranking. AutomationBench uses the public split; AA-Omniscience is an index on the public split only, calculated by Reflection AI. BrowseComp and DeepSearchQA use context management. AA-LCR has no stated version. These settings remain separate from other source fields.

Metric
CriPT AA
Published
5 October 2026
Effort
Not reported separately
Tasks / attempts
Unknown / Unknown
Confidence interval
Not reported
Build a reproduction plan →
GPQA Diamond90.5%self
Details

Reflection AI’s four capability tables, read 6 October 2026. All 21 rows and eight model columns are transcribed; NR means not reported and remains blank. The source’s Qwen 3.8 Max column uses this catalogue’s Qwen3.8 Max spelling. Beam results are publisher reports; Reflection AI cites Artificial Analysis and Datacurve for comparison-model evaluations. The announcement does not specify a common harness, effort or trial count for every row. Ranks are derived within each published row from reported values only, with ties sharing a rank; they do not establish an independent or overall model ranking. AutomationBench uses the public split; AA-Omniscience is an index on the public split only, calculated by Reflection AI. BrowseComp and DeepSearchQA use context management. AA-LCR has no stated version. These settings remain separate from other source fields.

Metric
GPQA Diamond
Published
5 October 2026
Effort
Not reported separately
Tasks / attempts
Unknown / Unknown
Confidence interval
Not reported
Build a reproduction plan →

API & pricing

Open one section for prompt tiers, runtime rates, endpoint measurements and effort configurations.

Published pricing, runtime and evaluator details
Access and weights

Waitlist preview while red-teaming continues. Apache 2.0 weights, a technical report, model card and developer artifacts are planned for October 2026; no public weight download is available in the checked announcement.

Read source ↗
Training

Reflection AI reports 23.8T pretraining tokens and over 100M reinforcement-learning rollouts across 10.5K NVIDIA GB300 GPUs over four weeks, with rollouts up to 256K tokens.

Read source ↗
Context and output

The trained context is 1M tokens. The beta API accepts 262,144 prompt and output tokens combined, with up to 131,072 output tokens including reasoning. Beta limits may change.

Read source ↗
API model

Beam-501B-A23B, created 5 October 2026. Knowledge cutoff: 30 June 2026.

Read source ↗
Reasoning

Always enabled: low, medium (default), high, xhigh or max. The text-only model supports tool calling and structured outputs; reasoning tokens count toward the completion limit.

Read source ↗
Evaluations

21 benchmarks in Reflection AI’s release tables. No independent Beam result was found in the checked Vals AI, Artificial Analysis, Datacurve or Scale AI leaderboards.

Read source ↗

Documentation read 6 October 2026.

Hosted runtime: Reflection AI
Context
262,144 tokens (prompt + output)
Modalities
Text
Capabilities
tools, thinking, cloud
Parameters / active
501B total parameters (MoE) / 23B active parameters per token

Public runtime prices are not catalogued.

Waitlist beta. Context includes prompt and output; reasoning counts toward the output cap. Public list prices are not published.

Runtime source · read 6 October 2026 ↗

Original sources