Anthropic Launches Claude Sonnet 5: Mid-Tier Model Challenges Flagship Opus on Agentic Tasks

Anthropic released Claude Sonnet 5 on June 30, 2026, positioning the mid-tier model as a cheaper alternative to its flagship Opus series for agentic tasks. The model is available immediately across all Anthropic plans and on Amazon Bedrock.

Benchmarks published by Anthropic show Sonnet 5 beating its predecessor Sonnet 4.6 across every tested category and matching or exceeding the larger Opus 4.8 on agentic coding, computer use, and knowledge work. On the knowledge work benchmark GDPval-AA v2, Sonnet 5 scored 1,618 points versus Opus 4.8 at 1,615 points.

Agentic Performance Jumps

Anthropic has built Sonnet 5 to act more independently than prior Sonnet models. It can create plans, use browsers, and operate terminals without step-by-step human guidance. The company reports that early-access partners consistently observed stronger agentic behavior during testing.

• On SWE-bench Pro (agentic coding), Sonnet 5 hit 63.2 percent, up from Sonnet 4.6 at 58.1 percent. Opus 4.8 sits at 69.2 percent.
• On Terminal-Bench 2.1, Sonnet 5 reached 80.4 percent versus 67.0 percent for Sonnet 4.6.
• On OSWorld-Verified (computer use), Sonnet 5 posted 81.2 percent, compared to 78.5 percent for its predecessor.
• On Humanity's Last Exam (multidisciplinary reasoning with tools), Sonnet 5 scored 57.4 percent, nearly matching Opus 4.8 at 57.9 percent.
• On BrowseComp (agentic search), Sonnet 5 outperformed Sonnet 4.6 at every effort level while offering cheaper entry points.

Pricing and Safety Considerations

Anthropic set introductory pricing for Sonnet 5 at $2 per million input tokens and $10 per million output tokens through August 31, 2026. After that date, prices rise to $3 and $15 respectively, matching previous Sonnet pricing. The model uses a one-million-token context window with a training cutoff of January 2026.

The launch comes against the backdrop of the US government blocking Anthropic's two most capable models, Mythos 5 and Fable 5, over cybersecurity concerns. Anthropic was quick to note that Sonnet 5 was not trained on cybersecurity tasks and that its scores for risky capabilities, such as writing software exploits, sit far below both Opus 4.8 and Mythos 5. In Firefox 147 exploit testing, Sonnet 5 showed a partial control rate of 13.2 percent, slightly higher than Sonnet 4.6 but well below the blocked models.

Anthropic has enabled cyber safeguards by default on Sonnet 5, matching the protections on Opus 4.7 and 4.8. The company assesses the overall cybersecurity risk from Sonnet 5 as low. The model also shows reductions in hallucinations and sycophantic behavior compared to Sonnet 4.6.

Developers can access Sonnet 5 via Claude Code, the Claude Platform, and the API using the identifier "claude-sonnet-5." The model replaces Sonnet 4.6 as the default for Free and Pro users and is available to Max, Team, and Enterprise subscribers.