At launch, Sonnet 5 is priced at an introductory rate of $2 per million input tokens and $10 per million output tokens through August 31, 2026. After that period, input pricing increases to $3 per million tokens and output pricing rises to $15 per million tokens. Anthropic said developers can access the model through the Claude API as claude-sonnet-5.
Performance improvements extend beyond coding. Anthropic says Sonnet 5 delivers stronger reasoning, tool use, knowledge work, and agentic capabilities than Sonnet 4.6 while approaching Opus 4.8 on several key evaluations. On one agentic coding benchmark highlighted by the company, Sonnet 5 scored 63.2%, compared with 69.2% for Opus 4.8 and 58.1% for Sonnet 4.6.
Anthropic argues that the release changes the economics of agentic AI by bringing higher-end autonomous capabilities into a lower-priced model tier. Rather than positioning Opus as the only option for complex workflows, the company recommends developers balance cost and accuracy by selecting between Sonnet 5 and Opus 4.8 depending on the task.
“It can make plans, use tools like browsers and terminals, and run autonomously at a level that, just a few months ago, required larger and more expensive models,” Anthropic said in its announcement.
The company also highlighted improvements in reliability during extended workflows. According to feedback from early access partners, Sonnet 5 is more likely to complete complex tasks without stopping midway and is better at checking its own work without explicit instructions.
Alongside the capability gains, Anthropic says Sonnet 5 improves on safety compared with Sonnet 4.6. Internal evaluations found lower rates of hallucinations, sycophancy, cooperation with misuse, deception, and successful prompt-injection attacks. The model also showed stronger refusal behavior when presented with malicious requests.
Anthropic emphasized that Sonnet 5 was not specifically trained for cybersecurity work and remains substantially less capable than its Opus models on evaluations involving dangerous cyber tasks. In testing involving software exploits for Firefox vulnerabilities, the company said Sonnet 5 was never able to produce a fully working exploit, although it achieved slightly more partial successes than Sonnet 4.6. Because of those improvements, Anthropic is enabling cyber safeguards by default to detect and block dangerous cybersecurity activity in real time.
Despite the advances, Anthropic says Opus 4.8 remains its preferred model for tasks requiring the highest levels of accuracy. “Opus 4.8 is still the model of choice for higher accuracy on these tasks, but Sonnet 5 provides developers with lower-priced options that are of much higher quality than what was previously available.”
The launch expands Anthropic’s strategy of bringing more advanced agentic AI capabilities to a broader audience instead of reserving them for its highest-end models. By making Sonnet 5 the default model across its consumer plans and pricing it below Opus, the company is positioning it as a lower-cost option for developers and businesses building AI-powered workflows.
This analysis is based on reporting from Anthropic.
Image courtesy of Anthropic.
This article was generated with AI assistance and reviewed for accuracy and quality.
