Introducing Claude Sonnet 5: The Most Agentic Sonnet Model Yet

For many developers, the agentic AI era began with Sonnet-class models: Claude Sonnet 3.5, 3.6, and 3.7 were the first models that showed impressive skills in coding and tool use. More recently, though, the clearest gains in agentic capabilities have been in our Opus-class models. Sonnet 5 narrows the gap: its performance is close to that of Opus 4.8, but at lower prices.

It is a substantial improvement over its predecessor, Sonnet 4.6, on important aspects of agentic performance like reasoning, tool use, coding, and knowledge work. Safety assessments also found that Sonnet 5 shows an overall lower rate of undesirable behaviors than Sonnet 4.6, making it safer to use in agentic contexts.

The benchmark evidence is specific. On BrowseComp, which tests complex agentic search tasks, Sonnet 5 at higher effort levels matches Opus 4.8, while at lower effort levels it still outperforms Sonnet 4.6. For reference, Sonnet 4.6 scored 34.6 percent on Humanity's Last Exam without tools and 46.8 percent with tools; Sonnet 5 improves on both numbers. On OSWorld-Verified, which measures computer use accuracy, Sonnet 4.6 now carries an updated reference score of 78.5 percent, and Sonnet 5 shows comparable gains relative to that baseline. Across these evaluations, Sonnet 5 represents what Anthropic describes as a strict improvement over Sonnet 4.6, not just in narrow cases.

Early access partners report that Sonnet 5 is much more agentic than its predecessors. Testers described how it finishes complex tasks where previous Sonnet models would stop short, how it checks its own output without explicitly being asked, and how it does all this agentic work at an attractive price point.

The partner feedback covered a range of industries. Legal teams found that Sonnet 5 handles multi-document research and analysis tasks that previously required human review to catch errors. Insurance teams reported end-to-end workflow automation, from intake through documentation, with less manual intervention than earlier Sonnet models required. Operations teams described Salesforce workflows that chain together record updates, approval triggers, and external notifications without stopping to ask for clarification at each step. These are sustained, multi-tool tasks, not single-shot prompts, and Sonnet 5 completes them at a higher rate than Sonnet 4.6.

At the Claude Platform, Sonnet 5 launches with introductory pricing of $2 per million input tokens and $10 per million output tokens through August 31, 2026. After that date, it will be priced at $3 per million input tokens and $15 per million output tokens. Developers can use the model ID claude-sonnet-5 via the Claude API today.

Safety audits show that Sonnet 5 produces fewer undesirable behaviors than Sonnet 4.6 overall. It is better at refusing malicious requests, more resistant to prompt injection attacks, and produces lower rates of hallucination and sycophancy. On Anthropic's automated behavioral audit, Sonnet 5 is assessed as safer than Sonnet 4.6 for agentic deployment, though it shows higher misaligned behavior rates than Opus 4.8 or Claude Mythos Preview. For teams deploying at scale, that distinction matters: Sonnet 5 is the right choice for most agentic workloads, while tasks requiring the strictest safety profile should still use Opus 4.8 or higher.

On cybersecurity specifically, Sonnet 5 scored 0.0 percent on tests involving development of working exploits for Firefox vulnerabilities, maintaining the safety floor Anthropic sets for Sonnet-class models. Cyber safeguards are enabled by default, consistent with Opus 4.7 and 4.8. The model does show slightly higher partial success on some security tasks than Sonnet 4.6, which is consistent with its broader capability improvements rather than a safety regression.

Sonnet 5 uses an updated tokenizer, similar to the one introduced with Opus 4.7. This change affects how the model processes text: the same input may now map to between 1.0 and 1.35 times more tokens depending on content type. Technical or code-heavy content tends to see the larger end of that range. Teams comparing costs against earlier Sonnet models should recalibrate their estimates against Sonnet 5 actual token consumption rather than assuming a direct swap. Pricing benchmarks built on Sonnet 4.6 usage data will otherwise undercount the true cost, particularly for developer and engineering workloads.

Claude Sonnet 5 is available on all Claude plans: Free, Pro, Max, Team, and Enterprise. It is the default model for Free and Pro users. Enterprise teams on partner platforms can access it through AWS Bedrock, Microsoft Azure Foundry, and Google Cloud Vertex AI. Rate limits have been increased across Chat, Cowork, Claude Code, and the Claude API to account for higher token usage at elevated effort levels. Operators building on the Claude Platform should review their rate limit allocations before routing high-volume workloads to Sonnet 5 at maximum effort.

The previous best Sonnet model (Sonnet 4.6) fell well short of Opus 4.8. Now Sonnet 5 and Opus 4.8 cover a single range, with Sonnet 5 offering impressive capabilities at a lower cost and Opus 4.8 offering greater accuracy at a higher price. Between Sonnet 5 and Opus 4.8, users can adjust the effort level to find the right balance of cost and performance.

Opus 4.8 is priced at $5 per million input tokens and $25 per million output tokens after introductory periods, roughly 1.7 times the standard Sonnet 5 rate per token. For workflows where Sonnet 5 at high effort matches Opus 4.8, that pricing gap is significant at volume. For workflows requiring Opus-level accuracy on every response regardless of cost, Opus 4.8 remains the better fit. Teams with mixed workloads can route task types to the model that fits their accuracy and cost requirements, using the effort level setting to fine-tune where each model sits on the cost-performance curve.