Atlas

Models

← All models

Opus 4.5

Anthropic · 2025-11-24

Claude Opus 4.5 is Anthropic's Opus-tier hybrid-reasoning model for coding, agents, computer use, and enterprise workflows, released with state-of-the-art performance on real-world software-engineering evaluations. It introduced the effort parameter with low, medium, and high levels and cut API pricing to $5/$25 per million input/output tokens, one-third of Opus 4/4.1. Anthropic documents a 200K-token hosted context window, although its system card reports one internal FinanceAgent evaluation at 1M context.

Benchmark scores

BenchmarkScore
A-Fantasia35.3
A-Fantasia Backwards Spelling16.0
A-Fantasia Chess22.0
A-Fantasia Cube Rotation68.0
AA-LCR74.0
AA-Omniscience Index13.3
Agent Red Teaming k=100 (CAIS)42.5
AIME 2024 (avg@8)99.6
AIME 2024-202595.4
AIME 202592.8
AIME 2025 (avg@8)91.3
ALE-Bench1429
AlgoTune1.8
APEX (Mercor)7.7
APEX-Agents20.7
APEX-SWE38.7
APEX-v1-extended57.3
App-Bench67.5
ARC-AGI-180.0
ARC-AGI-237.6
Arena Code WebDev — Brand & Marketing1467
Arena Code WebDev — Consumer Product1460
Arena Code WebDev — Content Creation Tools1446
Arena Code WebDev — Data & Analytics1447
Arena Code WebDev — Gaming1482
Arena Code WebDev — HTML1481
Arena Code WebDev — React1455
Arena Code WebDev — Reference-Based Design1472
Arena Code WebDev — Simulations1464
Arena Document (Style Controlled)1464
Loading Atlas data…