Step 3.5 Flash 2603 — benchmark results

Provider: StepFun. Released 2026-03-01. Access: API.

Unified ELO 1701 ± 29, rank #279 of 1841 rated models, from 33 benchmark results.

Strongest benchmark results

BenchmarkScoreMetricPercentile
AA Humanity's Last Exam22.57Accuracy (%)81.9
AA TAU-2 Bench87.43Accuracy (%)81.4
AA IFBench66.53Accuracy (%)78.9
AA CritPt2.29Accuracy (%)77.5
AA Terminal-Bench Hard32.58Accuracy (%)76.6
AA GPQA Diamond82.63Accuracy (%)76.5
AA Omniscience - Software Engineering (SWE) - Rust62Accuracy (%)74.1
AA Omniscience - Science, Engineering & Mathematics34.4Accuracy (%)73.5
AA Omniscience - Software Engineering (SWE) - Kotlin30Accuracy (%)73.4
Artificial Analysis Intelligence Index26Intelligence Index69.6
AA Omniscience - Humanities & Social Sciences25.7Accuracy (%)68.6
AA Omniscience - Business21.5Accuracy (%)68.5

Interactive version: theaggregate.ai/model?slug=step-3-5-flash-2603 · How It Works · Data refreshed daily, snapshot 2026-07-25.