Correction — October 3, 2026: We removed unsupported hands-on demonstrations and the blanket recommendation against Max effort. Published benchmark scores are now attributed to Anthropic, and the API model identifier and migration guidance are corrected. EyesTech has not independently reproduced these evaluations.
Anthropic released Claude Sonnet 5.5 on September 28, 2026. Its launch evaluation reports substantial gains over Sonnet 5, including 70.6% on Terminal-Bench 4.0. For a developer, the decision is whether it completes the application’s own tasks at an acceptable cost, rather than whether one benchmark puts it above a more expensive model.
The benchmark depends on the task and setting
| Anthropic-reported evaluation | Sonnet 5.5 | Opus 5.5 |
|---|---|---|
| Terminal-Bench 4.0 | 70.6% | 66.4%, Xhigh effort |
| FrontierCode 1.1 Main | 46.2% Max; 52.1% Xhigh | 54.4% |
| CursorBench 4.0 | 55.5% | 57.8% |
These are vendor-published results, not EyesTech measurements. The launch page includes harness and deployment caveats. Its FrontierCode footnote explains that Max effort sometimes triggered additional reviews, timeouts or out-of-scope edits. That is evidence about this evaluation; it is not proof that Max is always harmful.
A repository maintenance task may reward a small, scoped patch. An open-ended investigation may benefit from longer reasoning. Changing effort can change both behaviour and cost, so a recommendation needs the task definition, scoring rule and allowed time. Comparing percentages from unrelated tests would hide those differences.
Calculate the bill from the request
The model documentation lists the API identifier as claude-sonnet-5-5, a one-million-token context window and a maximum output of 128,000 tokens. Standard pricing is $2 per million input tokens and $10 per million output tokens; cache reads are $0.20 per million. Cache writes have separate rates.
An illustrative request with 100,000 uncached input tokens and 5,000 billed output tokens costs 0.1 × $2 + 0.005 × $10 = $0.25, before other billable services. Four such attempts would cost $1.00. Neither figure is a measured task cost, and a real agent may make additional calls or use a different mix of cached and uncached input.
Anthropic reports lower cost per task than Sonnet 5 and large savings at some lower effort settings on selected evaluations. Those claims do not establish a tenfold saving for every workload. Measure total spend through a successful, accepted result, including retries, tool work and human repair time.
Migration requires more than changing the model name
The migration guide documents changes to thinking and request parameters. Applications previously disabling thinking need to review the new between_tools mode and supported effort levels. Sampling controls and tool-selection behaviour also need checking against the current API contract.
Before changing a production route, replay a small set of representative requests: a successful tool call, a tool failure, a long conversation, a structured response and an intentionally restricted action. Verify both the result and the response shape your application parses. This is an evaluation procedure, not a claim that EyesTech ran those tests.
The same component-versus-workflow distinction matters in voice applications: our MAI streaming speech analysis explains why a component’s published latency is not the latency of a complete agent turn. For Sonnet, benchmark quality and token price likewise need testing inside the complete workflow.
Sonnet 5.5 is a candidate for routine and well-scoped work. Choose effort and escalation rules from observed failures in your own evaluation, and retain a route to stronger reasoning when the extra cost buys a better accepted result.
Cover: AI-generated editorial illustration; not a product screenshot or evidence of a measured result.
