Anthropic Releases Sonnet 5.5: Same Price, 30% Less Cost

Anthropic launched the Claude Sonnet 5.5 on Monday (28) at the same price as Sonnet 5: $2 per million input tokens and $10 per million output tokens, with cache reads at $0.20. The main change lies in performance. In the Terminal-Bench 4.0, an agent programming test, the model scores 70.6%, compared to 10.3% of its predecessor, according to the company’s launch page.

The critical data for large-scale AI buyers is different. Anthropic states that Sonnet 5.5 generates responses over 30% faster and costs up to 30% less per task in their tests, as it completes work with fewer tool calls and fewer tokens. The list price remains the same; the effective cost has decreased.

Mid-tier, Close to Top

The comparison with Opus 5.5, the company’s more expensive model, explains this move. Opus 5.5 costs $4 per million input tokens and $20 per million output tokens, double that of Sonnet 5.5 at both ends, according to the official pricing table. In GDPval-AA v2.1, which measures knowledge work, Sonnet 5.5 scores 1,844 Elo points, compared to 1,846 for Opus 5.5 and 1,449 for Sonnet 5.

The gap is not negligible in all tests. In OSWorld 2.1, which measures computer usage, Sonnet 5.5 scores 80.1%, against 81.8% for Opus 5.5. In Humanity's Last Exam with tools, the difference is larger: 64.5% versus 67.7%. For long reasoning tasks, the top model still delivers something the mid-tier model does not, and Anthropic positions Sonnet 5.5 for "well-defined daily tasks, bug fixing, and creating documents, presentations, and spreadsheets."

For a CIO, the calculus shifts. If a document analysis workflow currently runs on Opus due to lack of alternatives, the mandatory test becomes running the same batch on Sonnet 5.5 and measuring cost per completed task, rather than cost per token.

What Customers Measured

Customer data released by Anthropic support this direction. Balyasny Asset Management reports superior performance in 2,441 finance tasks, with token consumption falling from 497,000 to 121,000 per response, a reduction of about 76%. Box states the model was 2.4 times faster than Sonnet 5 in sensitive finance and health data. Zendesk claims "tickets were processed 20% faster."

Daniel Vogel, COO of Epic Games, said that Sonnet 5.5 "met the same quality benchmark expected of a top-tier model" in a systems design audit. Launch testimonials are selected by the provider and reflect favorable cases; they serve as a test hypothesis, not as independent benchmarks.

Cybersecurity Safeguards from Opus

Sonnet 5.5 arrives with cybersecurity protections comparable to those of Opus 5.5. Requests classified as higher risk "will visibly revert to Sonnet 5," according to Anthropic, and defenders needing expanded access can enroll in the Cyber Verification Program. For offensive security teams and red teams, this means that part of the work will visibly migrate back to the previous model, and the pilot needs to measure this fallback rate before any migration.

The model is already available on Anthropic's platform with the identifier claude-sonnet-5-5, as well as on AWS, Google Cloud, and Microsoft Azure, eliminating the contract barrier for companies already consuming AI from one of these three hyperscalers.

The Impact on Service Contracts

The most direct effect is not in labs, but in those who sell development hours. In the U.S., companies like Box and Zendesk are already embedding the speed gain into their products; the end customer receives the benefit without renegotiating anything. In India, where TCS, Infosys, and Wipro operate some of the largest software delivery centers, a jump from 10.3% to 70.6% in agent programming in a mid-tier model pressures fixed-price contracts: the client knows production costs have dropped and will seek discounts upon renewal.

In Brazil, where consultancies like CI&T sell development squads to American clients based on cost and time zone, the arbitration advantage shrinks when the mid-tier model performs the same task for 30% less. The exposure is common to both countries: those pricing by the hour absorb cost decreases as revenue drops, while those pricing by results capture the margin.

Anthropic maintained the token price even as the cost per task fell. For corporate buyers, the list metric has ceased to be a good indicator of what is paid; contracts comparing suppliers by price per million tokens are measuring the wrong thing.