VulcanBench VulcanBench
Model results · Reflection

Reflection Beam

Every published VulcanBench measurement of Reflection Beam (Reflection), newest first: combined score, Code quality, runtime and tokens at every reasoning level Beam offers on Frontier v4, graded by deterministic hidden tests on real tasks. Measured as Beam-1.0-501B-A23B (Reflection Beam 1.0, 501B total and 23B active parameters) through Mirror CLI 0.1.1, Reflection’s own coding agent, an early beta, on Reflection’s API with analytics off. Its Code quality is judged by Muse Spark 1.3 and GPT-6.1 Sol, a different second judge from most columns, because Grok 4.6’s Cursor account was out of usage. Reflection publishes no Beam price, so there is no cost.

Suitev4
Reflection Beam across every effort level
In Mirror CLI at Low to Max: combined score 39.51, 44.79, 55.01, 64.72 and 63.92 over judged runs (n=11, 20, 23, 23, 22); 18.90 at Low and 38.95 at Medium counting runs that left no code as 0. Tasks passed 0, 0, 0, 4 and 6 of 23; 0.4 to 22.4 minutes and 0.09M to 17.29M tokens a task. Best level Extra-high, 63rd of 75 columns on the board, about 24 to 26 points below Claude Opus 5.5, GPT-6.1 Sol and GPT-6 Astra, whose columns have a different judge panel.
October 27, 2026 · 23 tasks · 5 effort levels · 115 runs of Reflection Beam · VulcanBench Frontier v4 · Code quality protocol v3.28
→
Next

How would Reflection Beam do on your stack?

Public results answer "which model is stronger on this suite." Whether Reflection Beam belongs in your routing table depends on your languages, your codebases, and your task mix: that gets measured, not guessed. Grading rules are in the methodology.

Optimize your model routing All reports