Sonnet 5.5 Runs 30% Faster at Unchanged $2/$10 Pricing, Trailing Opus 5.5 by Just Two Points

News Summary

Anthropic released Claude Sonnet 5.5 on September 28, 2026, positioning it as a significantly cheaper and faster work partner for everyday tasks such as coding and creating office documents. Coverage published at about 11:00 AM Pacific Time that day reports that the model generates output more than 30 percent faster than Sonnet 5, keeps the same per-token price, and can cost up to 30 percent less per task because it uses fewer tokens and tool calls.

What Anthropic Announced

Sonnet 5.5 is the second model in the Claude 5.5 family, arriving roughly three months after Sonnet 5. Anthropic describes it as an ideal assistant for everyday tasks, including coding and creating office documents. TechCrunch reports that it outperforms Opus 5.5 on some agentic coding tasks and is better suited to multi-agent deployments where cost adds up quickly. Anthropic also says a new Haiku model, Claude Haiku 5.5, will follow in the coming weeks, although no date was given.

Pricing and Speed

According to VentureBeat and The Decoder, list prices are unchanged from Sonnet 5: $2 per million input tokens and $10 per million output tokens, with cache reads at $0.20 per million tokens and cache writes at $2.50 per million tokens. The savings therefore come from efficiency rather than a price cut. Output generation is reported to be more than 30 percent faster, and the model needs fewer tokens and tool calls to finish a job. Several outlets note that at Low or Medium effort, Sonnet 5.5 can beat Sonnet 5's best score on some benchmarks for about a tenth of the cost per task.

Benchmark Results

Reported figures show Sonnet 5.5 close to Opus 5.5, Anthropic's flagship, on knowledge-work and coding tests, and far ahead of Sonnet 5.

• GDPval-AA (knowledge work): 1844, versus 1846 for Opus 5.5 and 1449 for Sonnet 5.
• OSWorld 2.1 (computer use): 80.1 percent, versus 81.8 percent for Opus 5.5 and 57 percent for Sonnet 5.
• Terminal-Bench 4.0 (agentic coding): 70.6 percent, versus 66.4 percent for Opus 5.5 and 10.3 percent for Sonnet 5.
• CursorBench 4.0 (coding): 55.5 percent, versus 57.8 percent for Opus 5.5.
• Chartography (visual chart reading): 61.6 percent, versus 64.4 percent for Opus 5.5 and 15.6 percent for Sonnet 5.
• FrontierCode 1.1: 52.1 percent at high effort, up from 42.4 percent for Sonnet 5.

The Decoder adds a caveat: at the Max effort setting, the model sometimes scores lower than at Xhigh on certain benchmarks, because a code-review function can trigger unnecessary sub-agent work.

Early Customer Feedback

Companies that tested the model reported efficiency gains in their own workflows, as quoted by VentureBeat. Box said the model was 2.4 times faster and used 12 percent fewer total tokens. Zendesk reported 20 percent faster ticket processing, and Slack saw 14 percent fewer output tokens. Lovable measured about one-third fewer tool calls and roughly half as many shell executions, while Base44 reported 3.6 iterations per build compared with 7.7 for Opus 5. These are vendor-reported results and have not been independently reproduced.

Safety and Cyber Safeguards

Sonnet 5.5 is the first Sonnet model to ship with cybersecurity safeguards, because its cyber capabilities are described as comparable to those of Opus 5. Reported scores include 46.1 percent on CyScenarioBench versus 0.7 percent for Sonnet 5. High-risk requests are rerouted to Sonnet 5, and new classifiers help block attempts to extract reasoning or distill the model. Qualified security professionals can request tiered access through Anthropic's Cyber Verification Program. Biology safeguards are unchanged from Sonnet 5.

Availability and Effort Settings

The model is available on the Claude Platform under the identifier claude-sonnet-5-5, as well as on Amazon Web Services, Google Cloud, and Microsoft Azure. Zero data retention is supported. Developers can choose Low, Medium, or High effort. Claude Code and the consumer apps default to Medium, while the Claude Platform defaults to High.

Why It Matters for Learners and Builders

For students and developers, the release illustrates a broader trend in AI engineering: progress is increasingly measured in cost per completed task, not just raw benchmark scores. A faster model that reaches its answer with fewer steps lowers the cost of running many AI agents in parallel, which makes automated coding, document drafting, and data analysis more practical for small teams worldwide. Readers should treat the numbers above as reported by Anthropic and partners and check them against their own workloads.