Back to the index
Anthropic · Claude
No. 05

Claude Sonnet 5

The most agentic Sonnet yet — near-Opus coding reliability at Sonnet cost.

BalancedClosed weightsText, Image, Pdf
Field notes
Context window
1Mtokens
Max output
128Ktokens
Input price
$3/ 1M tokens
Output price
$15/ 1M tokens
Cached input
$0.30/ 1M tokens
Modalities
Text, Image, Pdf
Knowledge cutoff
Jan 2026
Released
Jun 2026
Approx. speed
Overview

What this model is

Claude Sonnet 5 is Anthropic's newest balanced-tier model and the most agentic Sonnet to date, closing much of the gap to Opus on real coding and terminal/tool-use work while keeping Sonnet pricing. It leads its predecessor decisively and edges Opus 4.8 on Terminal-Bench, sustaining long agent loops reliably — the new default production workhorse for most agentic and coding workloads. It pairs adaptive thinking with a 1M-token context and high-resolution vision. Introductory pricing of $2/$10 per 1M tokens runs through 2026-08-31.

Strengths

  • Near-Opus agentic-coding reliability at Sonnet cost
  • Best-in-Sonnet terminal and tool-use performance
  • 1M-token context with adaptive thinking
  • High-resolution vision
  • Excellent price-to-capability for production

Trade-offs

  • Opus 4.8 and Fable 5 still lead the hardest long-horizon jobs
  • 128K output cap
  • No audio or video input
  • New tokenizer counts ~30% more tokens than Sonnet 4.6

Best for

  • Production coding agents and Claude Code workflows
  • Terminal and tool-use automation
  • High-volume developer workflows
  • Everyday long-context work

Not ideal for

  • The very hardest autonomous coding (prefer Opus 4.8)
  • Ultra-low-cost bulk classification
Capabilities

Capability profile

Normalized 0–100 scores, comparable across the whole catalog.

Reasoning92Coding92Math90Writing92Knowledge91Speed70Agentic91Vision88Multilingual90Long Context94
Reasoning
92
Coding
92
Math
90
Writing
92
Knowledge
91
Speed
70
Agentic
91
Vision
88
Multilingual
90
Long Context
94
Benchmarks

How it scores

Public benchmark results, with independent third-party results where available. Bars normalize percentages to 100 and Elo ratings to a 1500 ceiling.

Independent index
Artificial Analysis Intelligence Index
Composite of ~9–10 independent evals · Artificial Analysis
53/ 100
BenchmarkResult
Artificial Analysis Intelligence Indexindependent
GeneralArtificial Analysis
53%
SWE-bench Proindependent
CodingSWE-bench Pro
63.2%
GPQA Diamondindependent
ReasoningIndependent aggregators
91.1%
AIME 2025
MathVendor-reported
100%
Alternatives

Comparable models

Models in a roughly similar class — worth weighing against this one.