Breaking

Anthropic Releases Claude Sonnet 5, Debuts 6th on Agent Arena

July 7, 2026 at 09:25 EDT

Foundation Models
AI Agents
Software Dev & Coding

Anthropic on June 30, 2026 released Claude Sonnet 5, which it calls its "most agentic Sonnet" model. On Agent Arena, which evaluates models on long-horizon tasks performed by real users, the Thinking variant debuted at No. 6 overall.

July 7, 2026 · Agent Arena

Claude Sonnet 5 Debuts at No. 6 on the Agent Arena Leaderboard

Anthropic's "most agentic Sonnet model yet" lands sixth overall on real-world long-horizon tasks — ranked live across 1,092,164 sessions and 33 models, aiming for near-Opus capability at Sonnet-tier pricing.

Net Improvement (overall) — 7.38%

Confirmed Success — 12.24%

1.14%

Tool Hallucination — low & stable

Net Improvement — Top of the Leaderboard

Anthropic models lead outright; Sonnet 5 sits sixth, closing on the Opus line.

14.07%

Claude Fable 5
#1

9.90%

Opus 4.8
#2

7.38%

Sonnet 5
#6

An Uneven Profile — Sonnet 5 by Metric

Strong on task completion, weaker on steerability and recovery.

Confirmed Success #2

12.24%

Praise vs Complaint ~#5

13.43%

Bash Recovery ~#9

9.15%

Steerability ~#15

0.95%

Praise

"A powerful execution layer for multi-step software engineering" — verifies its own output, sits on the price-performance Pareto frontier, and works as a cheap sub-agent alternative to Opus 4.8.

Criticism

At high effort it can cost more and perform worse than Opus 4.8. A new tokenizer raises token use ~1.0–1.35×, and some report weaker context retention — "not an Opus replacement."

Pricing & Specs

$2 / $10

per M input / output tokens (intro, through Aug 31 → then $3 / $15)

token context window · 128K max output

Default

model for Free/Pro · on Max, Team & Enterprise

Continue reading

The rest of this article is for AI News Blitz readers. Choose an option below to keep reading.