Anthropic released Claude Sonnet 5 on June 30, positioning it as the most agentic Sonnet-class model to date. The model closes the performance gap with Opus 4.8 on coding and autonomous task execution while maintaining Sonnet-tier pricing, according to Anthropic’s announcement.
Performance Gains
The largest improvements over Claude Sonnet 4.6 are concentrated in coding and agentic tasks. On BrowseComp (an agentic search evaluation) and OSWorld-Verified (a computer use benchmark), Sonnet 5 strictly outperforms Sonnet 4.6 across all effort levels and matches Opus 4.8 on some tasks, per Anthropic’s cost-performance analysis.
Anthropic describes the shift in explicit terms: “just a few months ago,” the agentic performance Sonnet 5 delivers “required larger and more expensive models.” Early access partners reported that Sonnet 5 finishes complex multi-step tasks where previous Sonnet models stopped short and checks its own output without being prompted to do so.
Sonnet 5 is now the default model for Free and Pro plan users on claude.ai, and is available in Claude Code and on the Claude Platform API.
New Tokenizer Changes Cost Math
Sonnet 5 introduces a new tokenizer that produces approximately 30 percent more tokens for the same input text, according to Anthropic’s documentation. Per-token pricing is unchanged, but the same request costs more in absolute terms because the tokenizer generates more tokens per word.
The practical effects for agent builders: context windows hold less text (1M tokens covers less content), output limits tuned for Sonnet 4.6 may truncate on Sonnet 5, and token counts measured against earlier models are no longer valid. Anthropic recommends recounting prompts and revisiting max_tokens budgets before migrating.
Three Behavior Changes
Sonnet 5 ships with three breaking API changes relative to Sonnet 4.6, per the documentation. Adaptive thinking is on by default (previously opt-in). Manual extended thinking with budget_tokens returns a 400 error (previously deprecated). Setting temperature, top_p, or top_k to non-default values returns a 400 error, a constraint previously introduced on Opus 4.7.
For teams running agents on Sonnet 4.6, the adaptive thinking change is the most consequential. Requests that ran without thinking now run with it, consuming additional tokens and requiring higher max_tokens limits.
Cybersecurity Safeguards
Sonnet 5 is the first Sonnet-tier model with real-time cybersecurity safeguards. Requests involving prohibited or high-risk cybersecurity topics may be refused, with refusals returning as HTTP 200 responses with stop_reason set to “refusal” rather than error codes, per Anthropic’s documentation. For security-focused agent workflows, this means adding refusal handling logic that was previously only necessary for Fable-class models.
Pricing
Introductory pricing is $2 per million input tokens and $10 per million output tokens through August 31, 2026. Standard pricing of $3/$15 takes effect after that, according to Anthropic. Combined with the 30 percent tokenizer increase, the effective per-request cost at standard pricing will be roughly 30 percent higher than equivalent Sonnet 4.6 requests for the same text.
Migration Tradeoff
The upgrade path is a model ID swap from claude-sonnet-4-6 to claude-sonnet-5. But the tokenizer change means the migration is not cost-neutral. Teams running high-volume agent workloads on Sonnet 4.6 need to benchmark their specific workloads before switching, because the performance gains come bundled with a higher token bill.