Anthropic

Claude Sonnet 4.5

Multimodal
#2τ-bench
#3OSWorld

by Anthropic

+
+
+
+
About

Claude Sonnet 4.5, released by Anthropic in September 2025, was built specifically around long-duration agentic work — coding, computer use, and autonomous tasks that run in loops, use tools, and execute over extended sessions. Anthropic positioned it as their most capable model for building complex agents, reflecting a shift toward designing Claude iterations primarily for autonomous agent use cases rather than conversational chat.

+
+
+
+
Pricing Range
Input (per 1M)$3.00 -$3.00
Output (per 1M)$15.00 -$15.00
Providers3
+
+
+
+
Timeline
ReleasedSep 29, 2025
Knowledge CutoffJan 1, 2025
+
+
+
+
Specifications
Capabilities
Multimodal
+
+
+
+
License & Family
License
Proprietary
Performance Overview
Performance metrics and category breakdown

Overall Performance

9 benchmarks
Average Score
44.6%
Best Score
84.7%
High Performers (80%+)
1

Performance Metrics

Max Context Window
264.0K

Top Categories

Agents
62.9%
Finance
54.5%
Coding
49.0%
Tool Use
43.8%
Reasoning
17.7%
+
+
+
+
All Benchmark Results for Claude Sonnet 4.5
Complete list of benchmark scores with detailed information
τ-bench
Agents
84.70
84.7%
Unverified
OSWorld
Agents
61.40
61.4%
Unverified
Finance Agent
Finance
54.50
54.5%
Unverified
Terminal Bench 2.0
Coding
51.00
51.0%
Unverified
SWE-rebench
Coding
47.10
47.1%
Unverified
MCP-Atlas
Tool Use
43.80
43.8%
Unverified
GDPVal
Agents
42.50
42.5%
Unverified
Humanity's Last Exam
Reasoning
17.70
17.7%
Unverified
MMMU
Multimodal
-1.00
-1.0%
Unverified