Anthropic Releases Claude Sonnet 5.5, 30% Faster at Unchanged Price

Anthropic released Claude Sonnet 5.5 on September 28, 2026, describing it as more than 30% faster and up to 30% cheaper per task than Sonnet 5, with per-token pricing unchanged at $2 per million input and $10 per million output tokens, half the rate of Opus 5.5. Per-task savings come from efficiency gains, fewer tokens and tool calls, rather than price cuts. An adjustable effort setting lets users trade reasoning depth against cost.

On Terminal-Bench 4.0, Sonnet 5.5 scored 70.6% compared with 10.3% for Sonnet 5 and 66.4% for Opus 5.5. On AutomationBench it reached 44.7%, beating Opus 5.5 by 2.2 points and GPT-6 Sol by 12.7 points. On GDPval-AA it scored 1844 Elo versus 1846 for Opus 5.5. One analysis calculated Sonnet 5.5 achieves 98% of Opus 5.5's benchmark scores, and at 1,000 bug-fix agent runs per month costs $420 versus $1,200 for Opus 5.5, with batch API cutting Sonnet 5.5's cost to $210.

Post-launch agentic testing showed Sonnet 5.5 fixed a multi-bug Python project in three tool calls where Sonnet 5 needed twelve. Real-world deployment data cited by one source showed Zendesk processed support tickets 20% faster, Slack used 14% fewer output tokens, and Balyasny cut finance task token usage from 497,000 to 121,000.

Anthropic says Sonnet 5.5 does not advance its absolute capability frontier, with Opus 5.5 remaining stronger at complex, open-ended work requiring sustained judgment. The company said benchmark scores capture only one facet of a model's capabilities. Sonnet 5.5 is positioned for well-scoped tasks such as bug fixes, ticket handling, slides, spreadsheets, and routine multi-file edits.

Sonnet 5.5 is now the model powering the free tier on claude.ai. One observer noted that ChatGPT's free tier uses Luna 5.6, making Anthropic's free offering more capable by that comparison. The model is the first Sonnet to include cyber safeguards, anti-distillation classifiers, and expanded preserved thinking. A known bug causes maximum thinking effort to exhaust the 128,000-token budget without producing output.

The model is available on the Claude API, Amazon Bedrock, Google Cloud, and Microsoft Foundry. Developer guidance published after launch flagged that forced tool choice, using 'any' or 'tool' values, returns a 400 error on both 5.5 models, and that the old fixed budget_tokens setting is removed for Opus 5.5, with depth now controlled via output_config.effort. That guidance also noted that hidden thinking tokens are billed as output, meaning a terse final answer alone will not solve over-spending if reasoning depth is high.

Anthropic said enterprise customers account for about 80% of its business, with clients including Salesforce, Databricks, Goldman Sachs, and Novo Nordisk. A third model, Haiku 5.5, designed for fast high-volume work, is expected in the coming weeks. Reuters reported the release came as CEO Dario Amodei called on the global AI community to slow the pace of releasing new capabilities over safety concerns.