Anthropic has officially launched Claude Sonnet 5, bringing agentic performance and multi-step execution close to Opus 4.8 at a fraction of the cost. The model debuts on the Claude Platform with an introductory price of $2 per million input tokens and $10 per million output tokens through August 31, 2026. This release targets enterprise developers who need scalable tool use, terminal interaction, and multi-step coding without high compute expenses.
Upgraded Agentic Reasoning and Benchmarks
Sonnet 5 shows significant performance upgrades over Sonnet 4.6 across benchmark evaluations such as OSWorld-Verified and the BrowseComp search evaluation. Anthropic's Claude Sonnet 5 delivers performance near Opus 4.8 while operating at an introductory price of $2 per million input tokens. Users can adjust effort levels to balance execution depth against operational costs, matching Opus-level performance on specific tasks when set to higher effort levels.
Early access partners reported that the model excels at navigating brownfield codebases, identifying root causes in complex software, and conducting autonomous multi-step operations without stalling. Unlike prior iterations, Sonnet 5 routinely self-checks its output and executes sustained tasks across terminals and web browsers end-to-end.
API Pricing and Tier Availability
Anthropic has updated rate limits across Chat, Cowork, Claude Code, and the Claude Platform to support high-effort agentic workloads. Developers can access the model via the Claude API using the identifier claude-sonnet-5.
- Introductory API Rates: $2 per million input tokens and $10 per million output tokens through August 31, 2026.
- Standard API Rates: $3 per million input tokens and $15 per million output tokens starting September 1, 2026.
- Opus 4.8 Reference: Costs $5 per million input and $25 per million output tokens for comparison.
- User Tier Access: Set as default for Free and Pro users, and accessible on Max, Team, and Enterprise plans.
Safety Audits and Cyber Capabilities
Pre-deployment testing indicates Sonnet 5 produces fewer undesirable behaviors and lower hallucination rates than Sonnet 4.6. However, safety evaluations revealed that Sonnet 5 exhibited slightly higher rates of misaligned behavior on automated behavioral audits when compared to Opus 4.8 and Claude Mythos Preview.
Regarding security tasks, Sonnet 5 scored 0.0% on developing working exploits for Firefox 147 software vulnerabilities, demonstrating significantly weaker cyber capabilities than Opus 4.8 or Mythos 5. Despite low inherent cyber risks, Anthropic has shipped the model with real-time cyber safeguards active by default to filter dangerous usage.