The Big Focus: Agentic Reliability
Rather than chasing a single headline benchmark, Anthropic has optimized Sonnet 5 for agentic reliability. In practice, this means the model is much better at:
- Executing longer task chains without losing context.
- Self-correcting on the fly when a tool call fails.
- Maintaining steady behavior during extended sessions inside environments like Claude Code or Cowork.
One of the most notable new features is the introduction of Effort Levels: low, medium, high, and xhigh (extra high). By dialing up the effort level, the model spends more tokens on internal reasoning. This significantly raises both the output quality and the cost, allowing developers to dynamically balance price and performance on a per-task basis.
Benchmark Breakdown: Closing the Gap with Opus 4.8
Anthropic’s published benchmarks show Sonnet 5 is a strict upgrade over Sonnet 4.6 across every tested category, aggressively closing the gap with the much more expensive Opus 4.8.
- SWE-bench Pro: Sonnet 5 scores an impressive 63.2%, a solid jump from Sonnet 4.6’s 58.1%. However, Opus 4.8 still maintains the crown here at 69.2%.
- OSWorld-Verified (Computer Use): Sonnet 5 hits 81.2%, up from 4.6’s 78.5%.
- Humanity's Last Exam (with tools): Sonnet 5 scores 57.4%, nearly matching Opus 4.8’s 57.9%.
- GDPval-AA v2 (Knowledge Work): In a surprising twist, Sonnet 5 actually edges ahead of the flagship model, scoring 1,618 against Opus 4.8’s 1,615.
API Pricing and Cost-Performance Tradeoffs
For developers, the cost-to-performance ratio is the most critical factor.
- Introductory Pricing (Through Aug 31, 2026): $2 per 1M input tokens / $10 per 1M output tokens.
- Standard Pricing (After Aug 31, 2026): $3 per 1M input tokens / $15 per 1M output tokens.
- Opus 4.8 Pricing: $5 per 1M input tokens / $25 per 1M output tokens.
The clearest value for Sonnet 5 lies in the low and medium effort levels. At these tiers, it delivers a level of agentic quality that previous Sonnet pricing simply couldn't buy. However, at the xhigh effort level, Sonnet 5 can actually end up costing more than Opus 4.8 for similar quality, due to the high volume of reasoning tokens consumed.
Real-World Workflows from Early Access
Early access partners have already put Sonnet 5 to work in production-grade scenarios:
- Autonomous Debugging: One tester asked Sonnet 5 to investigate a bug. In a single pass, the model wrote a reproducing test, implemented the fix, and confirmed the bug was resolved.
- End-to-End Automation: Zapier handed the model a multi-part job where it successfully updated Salesforce account tiers and autonomously sent a launch email to enterprise contacts.
How to Access Sonnet 5
Sonnet 5 is the new default model for Free and Pro plans, and is selectable for Max, Team, and Enterprise users. It is live in Claude Code, Cowork, and the Claude Platform.