Anthropic launches Claude Sonnet 5.5 with faster output and invisible watermarks

Anthropic launched Claude Sonnet 5.5 on Sept. 28, positioning it as its most capable mid-tier model for routine work such as drafting and summarising documents, spreadsheet tasks, coding fixes, interface polishing and creating slides. The company said the model generates output about 30% faster than Claude Sonnet 5, making it the fastest Sonnet version so far. It also said Sonnet 5.5 generally needs fewer tokens for the same work, producing an effective 30% reduction in usage costs even though its listed per-token prices are unchanged. Those prices are $2 per million input tokens, $10 per million output tokens and $0.20 per million cache reads. The Verge also reported Anthropic’s claim that the new model is 30% cheaper to use than Sonnet 5, consistent with the company’s explanation that lower token consumption, rather than lower list rates, produces the saving. On Terminal-Bench 4.0, an agentic coding evaluation, Sonnet 5.5 scored 70.6%, compared with 10.3% for the previous model. It scored two points below Anthropic’s flagship Opus 5.5 on GDPval-AA, a test of real-world occupational tasks. Anthropic also said it was the first Sonnet model to complete Pokémon Red using only screenshots. The release followed the Sept. 22 debut of Claude Opus 5.5. Sonnet 5.5 also incorporates cybersecurity safeguards and fallbacks derived from work on Anthropic’s more capable Mythos-class models and Fable, with stricter protections intended to limit misuse. In testing by Zendesk, the company said Sonnet 5.5 made fewer incorrect decisions than the Claude models it uses in production and processed support tickets 20% faster. The result came from hundreds of use cases involving replies and escalation requests, according to Zendesk’s director of AI, Abhinay Kathuria. The model also carries Anthropic’s invisible text watermarking. The company says the embedded mark does not affect a passage’s quality or readability, but is intended to make it more likely that AI-generated text can be identified. Sonnet 5.5 supports watermarks on Anthropic’s first-party products and outputs through AWS, Google Cloud and Microsoft Foundry; where supported, generated files receive Content Credentials under the C2PA standard. As we reported in August, Anthropic began watermarking newer Claude output as part of its response to European Union transparency rules. The company has signed the EU AI Act’s Article 50(2) Code of Practice on Transparency of AI-Generated Content, and its markings apply globally wherever supported Claude models are offered. The legal transparency obligations took effect on Aug. 2, while participation in the related code is voluntary. The European Commission and the AI Board have said the code is an adequate voluntary means of demonstrating compliance; about 190 companies and organisations had signed it by the end of July. Anthropic says Claude models launched in the EU from Aug. 2 onward support machine-readable marking at launch. The markers have limits. Anthropic says a watermark is imperceptible, stays with copied-and-pasted text and may survive some edits, but a detected mark only signals that Claude may have generated or processed the content. Its absence does not show that text was not generated or processed by AI. Anthropic’s detection capability remains in private preview for eligible organisations and enterprises with compliance obligations, although it plans to broaden access to the API over time. Brookings notes that AI-generated text is harder to identify reliably than audio-visual material, in part because it offers fewer detectable artifacts. It adds that traditional text watermarks can be damaged or removed by small edits, and that no technically sound detection tool performs well across every relevant policy consideration. Anthropic said it plans to release Claude Haiku 5.5, its smallest and most efficiency-focused model for high-volume and cost-sensitive uses, in the coming weeks.