Claude Sonnet 5.5 is Anthropic’s efficiency-focused model for everyday professional work. Released on September 28, 2026, it is positioned below Opus 5.5 for the hardest open-ended problems, but it brings meaningful gains in coding, document creation, image understanding, and long-running tasks. The most practical changes are straightforward: Anthropic says it generates output more than 30% faster than Sonnet 5, keeps the same API token prices, and can cost up to 30% less per completed task because it often uses fewer tokens.
It is designed for people who want a capable model for well-defined work without paying for Opus on every request. Model choice is rarely about the highest benchmark score alone; latency, retries, token use, and human review can have a larger effect on the real cost of a workflow.
Claude Sonnet 5.5 at a Glance
| Detail | Claude Sonnet 5.5 |
|---|---|
| Release date | September 28, 2026 |
| API model name | claude-sonnet-5-5 |
| Standard input price | $2 per 1M tokens |
| Standard output price | $10 per 1M tokens |
| Cache read price | $0.20 per 1M tokens |
| Cache write price | $2.50 per 1M tokens |
| Best fit | Well-scoped coding and professional tasks |
| Availability | Claude products, Claude Platform, AWS, Google Cloud, and Microsoft Azure |
These figures come from Anthropic’s Sonnet 5.5 announcement. Cloud-provider pricing and access conditions can differ, so teams should check the platform they actually use before estimating production costs.
What Changed From Sonnet 5?
The update is strongest where a model must do work rather than simply answer a question. Anthropic highlights bug fixing, polished documents, presentations, spreadsheets, visual design, and longer tasks. Sonnet 5.5 also supports adjustable effort: lower settings prioritize speed and lower token use, while higher settings allow more reasoning and checking. Anthropic sets Medium as the default in Claude Code and its apps, while the Claude Platform defaults to High.
The headline speed and cost claims need careful wording. “30%+ faster” refers to output generation compared with Sonnet 5. “Up to 30% less” refers to Anthropic’s observed cost per task, not a discount on each token. The listed API price remains $2 per million input tokens and $10 per million output tokens, so actual savings depend on whether the new model finishes your work with shorter answers, fewer retries, or fewer tool calls.
Claude Sonnet 5.5 Benchmarks
Anthropic published several evaluations spanning agentic coding and knowledge work. They show a large improvement over Sonnet 5 in some environments and near-Opus results in others.
| Benchmark | Sonnet 5.5 | Sonnet 5 | Opus 5.5 | GPT-6 Sol |
|---|---|---|---|---|
| Terminal-Bench 4.0 | 70.6% | 10.3% | 66.4% | Not reported |
| FrontierCode 1.1 Main | 46.2% at Max | 42.4% | 54.4% | 49.3%; 52.1% at Xhigh |
| CursorBench 4.0 | 55.5% | 34.1% | 57.8% | Not reported |
| GDPval-AA v2.1 | 1844 | 1449 | 1846 | 1487 |
| AA-Briefcase v1.1 | 1811 | 1359 | 1822 | 1483 |
The table is useful, but it is not a universal ranking. Effort settings, agent harnesses, tools, safeguards, and scoring methods vary. Anthropic also notes that more reasoning did not always improve results: on FrontierCode, Sonnet 5.5 at Max sometimes made changes outside the requested scope. The best configuration produces an acceptable result with controlled edits, not necessarily the one that spends the most compute.
Coding and Agent Work
Sonnet 5.5’s 70.6% result on Terminal-Bench 4.0 is the most eye-catching number in the release, but the surrounding workflow matters more than the headline. For a developer, a useful coding model must understand the repository, make a targeted change, run appropriate checks, and stop before expanding the task. Speed is valuable only when the output remains easy to review.
That makes Sonnet 5.5 a sensible candidate for bounded work such as diagnosing a bug, implementing a clearly specified feature, reviewing a pull request, or refactoring a known component. Large migrations and ambiguous architectural decisions may still benefit from Opus 5.5, which Anthropic explicitly positions for sustained judgment. A fair evaluation should therefore use real repository tasks and measure correctness, unnecessary edits, elapsed time, total tokens, and human review time.
Documents, Spreadsheets, and Knowledge Work
The release is equally relevant outside software development. Anthropic describes Sonnet 5.5 as strong at creating documents, slides, and spreadsheets, while the GDPval-AA and AA-Briefcase results are intended to reflect professional knowledge work. Those scores do not prove that every report will be accurate, but they support testing the model on structured tasks with clear source material and a verifiable deliverable.
A practical workflow might begin with several reports, meeting notes, and a spreadsheet, then ask the model to identify agreements and contradictions, extract figures with source references, and draft an executive brief. The important safeguard is verification: numerical claims should be checked against the source, and consequential conclusions still need qualified human review. If the first challenge is organizing mixed source material, iWeaver’s AI Summarizer can help turn PDFs, documents, webpages, audio, images, and video into summaries or structured notes before deeper analysis.
Pricing in Context
Sonnet 5.5 and GPT-6 Sol share the same standard list price of $2 per million input tokens and $10 per million output tokens. Opus 5.5 costs $4 and $20 respectively. Those figures make Sonnet attractive for repeated, well-defined tasks, but a token-rate comparison is incomplete. Context size, caching, tool charges, reasoning effort, failed runs, and output length all affect the final bill.
For production evaluation, choose a representative set of tasks and record total task cost rather than price per token alone. A lower-priced model that requires repeated correction may cost more than expected, while a higher-priced model can be economical when it finishes a high-value task in one pass. The same test set should also be rerun after model updates, because provider behavior and serving configurations can change.
Is Claude Sonnet 5.5 Worth Using?
Sonnet 5.5 is worth testing if most of your work is clearly scoped and you value responsiveness, predictable API pricing, and strong coding or document performance. It is less obviously the right default when the problem is ambiguous, the cost of a subtle error is high, or the task depends on judgment sustained across many stages; those cases are closer to Opus 5.5’s stated role.
The strongest conclusion is not that Sonnet 5.5 “beats” every alternative. It is that Anthropic has narrowed the gap between its everyday and premium tiers while preserving a meaningful price difference. Use the benchmarks to decide what deserves a trial, then choose based on your own sources, prompts, review standards, and total task cost. For a direct tier decision, see Claude Sonnet 5.5 vs Opus 5.5; for setup ideas, continue with How to Use Claude Sonnet 5.5.