Atlas

Models

← All models

Claude Sonnet 5.5

Anthropic · 2026-09-28

Claude Sonnet 5.5 is Anthropic's September 28, 2026 model for coding and professional work. Adaptive thinking is on by default with high effort; low, medium, high, xhigh, and max are supported. The between_tools setting disables up-front thinking at high effort or below. Provider fallbacks can route declined requests to Sonnet 5. Its published June 2026 knowledge cutoff has no day precision.

Benchmark scores

BenchmarkScore
AA-3D v1.0 Elo1199
AA-AnalystAgent (pass@1)69.5
AA-AnalystAgent (pass@5)78.8
AA-AnalystAgent (pass^5)57.5
AA-Briefcase v1.1 Elo1812
AA-LCR v1.179.7
AA-Omniscience Index32.3
AA-Omniscience v1.1 Accuracy51.5
AA-Omniscience v1.1 Hallucination Rate43.3
AA-Omniscience v1.1 Index30.5
ActiveVision41.2
ActiveVision with tools78.8
AdvancedIF83.4
AdvancedIF (rubric mean score)94.0
AIM Consistency94.5
AIM Grounding86.6
AIME 2025100.0
ALE-Bench2448
APEX (Mercor)17.5
APEX-Agents 1.183.7
APEX-Agents 1.1 Pass@175.5
APEX-SWE66.4
APEX-v1-extended53.9
ARC-AGI-1 Semi-Private96.5
ARC-AGI-283.9
ARC-AGI-3 (Semi-Private)72.4
Arena Alignment — Deceptive Completion9.5
Arena Alignment — False Attribution5.8
Arena Alignment — Unauthorized Action1.8
Arena Alignment Index79.5
Loading Atlas data…