Claude Sonnet 5: Anthropic's Most Agentic Mid-Tier Model

Published: June 30, 2026 Source: Anthropic Official Announcement

Anthropic launched Claude Sonnet 5 on June 30, 2026, positioning it as the "most agentic Sonnet model yet." The model represents a substantial upgrade over Sonnet 4.6, narrowing the capability gap to Opus 4.8 while maintaining Sonnet-tier pricing.

Key Specifications

Attribute	Detail
Context Window	1,000,000 tokens
Maximum Output	128,000 tokens
Introductory Pricing	$2/M input, $10/M output (through Aug 31, 2026)
Standard Pricing	$3/M input, $15/M output (from Sep 1, 2026)
API Identifier	claude-sonnet-5
Tokenizer	Updated tokenizer (similar change introduced with Opus 4.7)
Default Model	Yes — for Free and Pro plans

Benchmark Performance

All benchmark figures sourced from the Claude Sonnet 5 System Card published by Anthropic.

Coding: SWE-Bench Pro

Model	Score
Opus 4.8	69.2%
Sonnet 5	63.2%
Sonnet 4.6	58.1%

Sonnet 5 achieves a 5.1-point improvement over Sonnet 4.6, closing much of the gap to Opus 4.8.

Terminal Work: Terminal-Bench 2.1

Model	Score
Sonnet 5	80.4%
Opus 4.8	74.6%
Sonnet 4.6	67.0%

Sonnet 5 surpasses Opus 4.8 on terminal tasks — the first benchmark in the suite where the mid-tier model outperforms the flagship.

Hard Reasoning: Humanity's Last Exam (with tools)

Model	Score
Opus 4.8	57.9%
Sonnet 5	57.4%
Sonnet 4.6	46.8%

Sonnet 5 is essentially tied with Opus 4.8 on hard reasoning with tool use, representing a 10.6-point jump over Sonnet 4.6.

Computer Use: OSWorld-Verified

Model	Score
Opus 4.8	83.4%
Sonnet 5	81.2%
Sonnet 4.6	78.5%

Sonnet 5 narrows the gap to Opus 4.8 to just 2.2 points on computer use tasks.

Agentic Search: BrowseComp

Model	Score
Sonnet 5	84.7%
Opus 4.8	83.4%
Sonnet 4.6	73.1%

Sonnet 5 leads on agentic search, outperforming both Opus 4.8 and Sonnet 4.6.

Knowledge Work: GDPval-AA v2

Model	Score
Sonnet 5	78.5%
Opus 4.8	77.2%
Sonnet 4.6	65.3%

Sonnet 5 exceeds Opus 4.8 on knowledge work tasks, marking the second benchmark where it leads the flagship.

Safety Assessments

Anthropic's pre-deployment safety evaluations found Sonnet 5 to be an overall improvement over Sonnet 4.6:

• Lower rate of undesirable behaviors than Sonnet 4.6 across the automated behavioral audit
• Better at refusing malicious requests and resisting prompt injection hijack attempts
• Lower rates of hallucination and sycophancy than Sonnet 4.6
• Substantially poorer cybersecurity capabilities than Opus 4.8 and Mythos 5 — Sonnet 5 scored 0.0% on developing working exploits for Firefox vulnerabilities (same as Sonnet 4.6)
• Cyber safeguards enabled by default — same safeguards as Opus 4.7 and 4.8, less strict than those on Fable 5

However, Sonnet 5 shows a slightly higher rate of partial success on cybersecurity tasks than Sonnet 4.6, attributed to improvements in general intelligence rather than specific training.

Effort Levels

Sonnet 5 supports adjustable effort levels (low through extra-high), enabling users to trade cost for performance:

• At medium effort, Sonnet 5 provides substantially improved cost efficiency
• At extra-high effort, Sonnet 5's performance on OSWorld-Verified and BrowseComp approaches Opus 4.8's medium-to-high setting
• Users can select between Sonnet 5 and Opus 4.8 at different effort levels to find the optimal cost-performance balance

Availability

Claude Sonnet 5 is available across all Anthropic platforms:

• Claude Chat — default model for Free and Pro plans
• Claude Cowork — available to all plan tiers
• Claude Code — available for agentic coding workflows
• Claude Platform (API) — available via claude-sonnet-5 identifier
• Max, Team, and Enterprise plans — available alongside other models

Significance

Sonnet 5 represents a structural shift in Anthropic's model lineup. For the first time, a mid-tier Sonnet model outperforms the flagship Opus on specific benchmarks (Terminal-Bench 2.1 and GDPval-AA v2), while matching or nearly matching Opus 4.8 across coding, reasoning, computer use, and agentic search.

The pricing strategy — introductory rates of $2/$10 per million tokens, rising to $3/$15 after August 31 — positions Sonnet 5 as the default choice for most agentic workloads, with Opus 4.8 reserved for tasks requiring the last few accuracy points.

The updated tokenizer (similar to the change introduced with Opus 4.7) increases token counts by roughly 1.0–1.35× depending on content type, which Anthropic offsets with the introductory pricing to make the transition roughly cost-neutral for users migrating from Sonnet 4.6.

References

1. Anthropic. (2026, June 30). Introducing Claude Sonnet 5. https://www.anthropic.com/news/claude-sonnet-5
2. Anthropic. (2026, June 30). Claude Sonnet 5 System Card. https://www.anthropic.com/claude-sonnet-5-system-card
3. Anthropic. (2026, June 30). Release Notes: Claude Sonnet 5 launch. https://docs.anthropic.com/en/release-notes/claude-apps
4. Anthropic. (2026). Claude Platform — Models Overview. https://platform.claude.com/docs/en/about-claude/models/overview
5. Anthropic. (2026). Real-time Cyber Safeguards on Claude. https://support.claude.com/en/articles/14604842-real-time-cyber-safeguards-on-claude