Claude Sonnet 5.5: Speed, Cost, and the New Enterprise Workhorse

What changed in Claude Sonnet 5.5

Anthropic released Claude Sonnet 5.5 on September 28, 2026 as the second model in its Claude 5.5 family. The company describes it as a faster, lower-cost complement to Opus 5.5: designed for well-scoped everyday tasks, bug fixing, and polished documents, slides, spreadsheets, and interfaces.

Anthropic reports that the model generates output more than 30% faster than Sonnet 5 and can cost up to 30% less per task because it often uses fewer tokens. The published list price remains $2 per million input tokens and $10 per million output tokens, with cache reads at $0.20 per million and cache writes at $2.50 per million.

This distinction between token price and task price is important. A model that completes work in fewer tool calls and shorter reasoning trajectories can reduce total operating cost without changing its rate card.

Coding gains point to a broader agentic improvement

Anthropic reports 70.6% on Terminal-Bench 4.0 for Sonnet 5.5, compared with 10.3% for Sonnet 5 under its published setup. On FrontierCode 1.1, the company reports a 10-point improvement at High effort over Sonnet 5 at the same setting. Provider-reported benchmark results are useful signals, but they are not substitutes for private evaluations on your repositories and tools.

The more operationally interesting claim is that early testers observed fewer steps, fewer failed tool calls, and better context gathering. Those characteristics influence whether an agent can complete work reliably inside a time and cost budget. They also affect usability: a system that reaches the right outcome with fewer interruptions creates less supervision overhead.

For engineering teams, sensible pilots include bounded bug fixing, dependency updates, test repair, code review, documentation maintenance, and migrations with strong automated checks. Autonomous production changes should remain behind branch protection, review, tests, and deployment controls.

The enterprise opportunity extends beyond software

Anthropic positions Sonnet 5.5 for professional documents, financial analysis, support workflows, research, and visual interpretation. Its launch materials report performance close to Opus 5.5 on GDPval-AA, a benchmark spanning real-world work across multiple occupations, and stronger chart recognition than Sonnet 5.

The practical implication is not that organizations should automate every document workflow. It is that repeatable knowledge work can be decomposed into controlled stages: retrieve approved sources, extract evidence, calculate or transform data, draft an output, verify it against a rubric, and route exceptions to a person.

That architecture makes quality measurable. It also allows different models or effort levels to handle different stages instead of paying for maximum reasoning throughout the entire process.

• Customer-support drafting with policy retrieval and escalation.
• Financial or operational analysis with source-linked evidence.
• Document production against controlled templates and rubrics.
• High-volume software maintenance with automated verification.

Effort controls should become part of workflow design

Claude's effort setting changes how long the model reasons and how much work it performs before returning an answer. Anthropic uses Medium by default in Claude apps and High on the Claude Platform. Lower effort can suit routine work; higher effort may improve difficult tasks but increases latency and token consumption.

Organizations should avoid selecting one effort level for every request. A classifier, policy rule, or initial model pass can route straightforward work to lower effort and reserve higher effort for ambiguous, high-value, or failed cases. The routing policy should be evaluated alongside the model itself because it directly affects cost and reliability.

What to test before migrating

Sonnet 5.5 is Anthropic's first Sonnet release with cyber safeguards and fallbacks similar to those used for more capable models. Anthropic also applies biology safeguards and anti-distillation protections, including preserved-thinking behavior tied to the account that created a conversation.

Before replacing an existing model, test normal requests, borderline security work, account and session transitions, long-running tool sequences, and structured-output edge cases. Review whether safeguards create new fallbacks or refusals in legitimate workflows, and document an escalation path rather than attempting to bypass controls.

The decision between Sonnet 5.5 and Opus 5.5 should be based on task distribution. Sonnet is the logical default for frequent, well-scoped work; Opus remains the comparison for complex, open-ended problems requiring sustained judgment. A mixed architecture will often outperform a single-model standard.