Claude Sonnet 5.5 costs exactly what Sonnet 5 cost per token: $2 per million input, $10 per million output. Whether it costs less per task depends on how hard you run it. Independent testing by Artificial Analysis shows a saving at Medium effort, $0.59 per index task against $1.00 for Sonnet 5, and a higher bill at Max, $7.60 against $5.09. Anthropic’s claim of up to 30% lower cost for most work describes the lower settings.
Anthropic released Sonnet 5.5 on September 28, 2026, six days after Opus 5.5, as the second model in the Claude 5.5 family, with Haiku 5.5 to follow in the coming weeks. Below: the price and cost-per-task data by effort, the benchmarks with Anthropic’s footnotes, the six changes that can break API integrations, the new safeguards, and when Sonnet 5.5 is the right pick over Opus 5.5, including the cases where it is not the cheaper one. We have not tested the model ourselves.
Claude Sonnet 5.5 at a glance
- Release date: September 28, 2026
- Position: faster, lower-cost complement to Opus 5.5 for well-scoped tasks, bug fixes, documents, slides, spreadsheets, and interface polish
- API model name: claude-sonnet-5-5
- Price: $2 input, $10 output, $0.20 cache reads, $2.50 cache writes per million tokens, unchanged from Sonnet 5
- Speed: Anthropic reports output more than 30% faster; Artificial Analysis measures 137 tokens per second at Max against 75 for Sonnet 5
- Default effort: Medium in the Claude apps and Claude Code, High on the Claude Platform
- Availability: Claude Platform, Amazon Web Services, Google Cloud, Microsoft Azure; zero data retention available
- Context window: 1 million tokens per Artificial Analysis; Anthropic’s page points to the model card
Price per token vs cost per task
| Price per million tokens | Claude Sonnet 5.5 | Claude Sonnet 5 | Claude Opus 5.5 |
|---|---|---|---|
| Input | $2.00 | $2.00 | $4.00 |
| Output | $10.00 | $10.00 | $20.00 |
| Cache reads | $0.20 | $0.20 | $0.20 |
| Cache writes, 5 minutes | $2.50 | $2.50 | $5.00 |
| Cache writes, 1 hour | $4.00 | $4.00 | $8.00 |
Source: Claude Platform pricing docs, September 29, 2026. The tokenizer is unchanged from Sonnet 5, so the same text produces the same token count; the difference is how many tokens the model spends reasoning and acting.
Anthropic’s “up to 30% less per task” comes from its own tests and from early customers: Box reports 12% fewer total tokens, Slack about 14% fewer output tokens, and one asset manager 121,000 tokens per answer against 497,000 for Sonnet 5. Artificial Analysis ran both models at every effort level on its ten-benchmark Intelligence Index, which is the only public dataset that lets you see how the saving moves with effort.
| Effort | Sonnet 5.5 score and cost per task | Sonnet 5 score and cost per task | Opus 5.5 score and cost per task |
|---|---|---|---|
| Low | 36, $0.41 | 24, $0.51 | 42, $0.55 |
| Medium | 41, $0.59 | 28, $1.00 | 51, $1.34 |
| High | 47, $1.08 | 32, $1.79 | 54, $1.82 |
| Xhigh | 52, $2.74 | 34, $2.87 | 56, $3.46 |
| Max | 56, $7.60 | 38, $5.09 | 58, $5.98 |
Source: Artificial Analysis release pages, index v4.3.2, September 29, 2026. Read down the Sonnet columns and the saving is real at Low through Xhigh, largest at Medium, the default in the Claude apps. At Max it reverses: Sonnet 5.5 spends 410 million output tokens across the suite against Sonnet 5’s 370 million and costs 49% more per task. The saving depends on effort and on the task mix, not on the model spending fewer tokens under all conditions.
Claude Sonnet 5.5 benchmark results
Anthropic’s table compares Sonnet 5.5 with Sonnet 5, Opus 5.5, and GPT-6 Sol. The footnotes change how several rows read.
| Benchmark | What it measures | Sonnet 5.5 | Sonnet 5 | Opus 5.5 | GPT-6 Sol |
|---|---|---|---|---|---|
| Terminal-Bench 4.0 | Agentic terminal coding | 70.6% | 10.3% | 66.4% (xhigh) | not reported |
| FrontierCode 1.1 Main | Merge-ready code changes | 46.2% (Max), 52.1% (Xhigh) | 42.4% | 54.4% | 49.3% |
| CursorBench 4.0 | Tasks from real Cursor sessions | 55.5% | 34.1% | 57.8% | not reported |
| GDPval-AA v2.1 | Real-world work across 44 occupations (Elo) | 1,844 | 1,449 | 1,846 | 1,487 |
| AA-Briefcase v1.1 | Multi-hour office work (Elo) | 1,811 | 1,359 | 1,822 | 1,483 |
| Humanity’s Last Exam, with tools | Multidisciplinary reasoning | 64.5% | 54.9% | 67.7% | not reported |
| OSWorld 2.1, partial | Computer use | 80.1% | 57.0% | 81.8% | not reported |
| Chartography, no tools | Chart reading | 61.6% | 15.6% | 64.4% | 53.6% |
Four footnotes from Anthropic. Opus 5.5’s Terminal-Bench figure is its xhigh score. Sonnet 5.5 scores lower at Max than at Xhigh on FrontierCode because at Max it more often ran a code-review skill that split work across subagents and produced out-of-scope edits or timeouts, which the benchmark penalizes. The two Artificial Analysis rows were run on a pre-release deployment with a structured-outputs bug that Anthropic says likely understated Sonnet 5.5 slightly and has since been fixed. And GPT-6 Sol’s GDPval, Briefcase, and Chartography scores may predate an OpenAI fix to image understanding; Terminal-Bench charts use GPT-5.6 Sol because OpenAI did not report GPT-6 Sol.
Two readings follow. Against Sonnet 5 the gains are large on every row, and Artificial Analysis independently scores Sonnet 5.5 at 56 against 38. Against Opus 5.5 the gap on GDPval-AA is two Elo points in a single evaluation without a published uncertainty interval, which is not evidence that the models are equivalent; Anthropic itself says that in its own and external testing Opus 5.5 remains clearly stronger at complex, open-ended work requiring sustained judgment. Early testers credit much of Sonnet 5.5’s efficiency to batching tool calls: Lovable reports a third fewer tool calls, Base44 3.6 iterations per app build against 7.7 for Opus 5. Those are vendor-published testimonials.
Migration checklist for API users
Anthropic’s what’s-new page lists five breaking changes for code running on Sonnet 5 and one change to the response shape. All six can affect production.
- Thinking off is now between_tools. Sending thinking disabled returns a 400 error; send type between_tools instead. It works at Low, Medium, and High effort only; at Xhigh or Max it returns 400, and effort cannot change mid-conversation with it.
- Forced tool use returns 400. tool_choice set to any or to a named tool is rejected. Keep auto, use strict tool use or structured outputs for schema-valid input, and prompt the model when a tool applies.
- Thinking blocks are bound to the model and the conversation. Sonnet 5.5 reads blocks from Sonnet 5, Opus 4.8, Haiku 4.5, and earlier, but not from Opus 5, Opus 5.5, Fable, or Mythos; no other model reads its blocks. For accounts created on or after August 31, 2026, replaying a block after the system prompt, tools, or an earlier message changed returns 400 by default. Keep histories append-only.
- The computer_20251124 tool is rejected on the Claude API and Google Cloud; computer use requires the computer_toolset_20260801 toolset.
- Some advisor pairings fail. With Sonnet 5.5 as executor, Opus 4.8, Opus 4.7, and Sonnet 5 advisors return 400; accepted advisors return their advice encrypted.
- Text between tool calls moves into thinking blocks. Notes longer than a sentence or two come back as progress-update blocks whose text is empty at the default display setting, so an application that streams those notes to users goes quiet between tool calls with no error. Set a display value that returns the text, or use between_tools.
Anthropic also says effort levels are recalibrated and recommends re-running your effort sweep rather than carrying a setting over. Thinking blocks are additionally tied to the account that produced them, which affects switching accounts mid-session in Claude Code.
Safety changes that affect real work
Sonnet 5.5 is the first Sonnet model with cybersecurity safeguards of the kind built for Opus 5.5, because Anthropic judges its cyber capabilities comparable to Opus 5’s. Routine bug finding and fixing is unaffected; higher-risk cybersecurity tasks visibly fall back to Sonnet 5, and Anthropic says defenders will soon be able to apply to an expanded Cyber Verification Program. Biology safeguards are unchanged from Sonnet 5. Refusals arrive as HTTP 200 with a stop reason and one of five categories: cyber, bio, frontier_llm, reasoning_extraction, or general_harms; server-side fallback retries only the cyber and frontier_llm categories on Sonnet 5.
Also new for the Sonnet tier: classifiers that block reasoning extraction, aimed at distillation attacks. On the automated behavioral audit of roughly 1,850 scenarios, Sonnet 5.5 matches or improves on Sonnet 5 on most measures and comes close to Opus 5.5 on containment; Anthropic repeats that no evaluation set catches every failure.
Sonnet 5.5 or Opus 5.5?
Anthropic’s own division is Sonnet 5.5 for well-scoped everyday tasks, bug fixes, and polished documents, slides, and spreadsheets, and Opus 5.5 for complex, open-ended work requiring sustained judgment. Half the per-token price does not settle the question by itself. At Max effort, Sonnet 5.5 reaches an index score of 56 for $7.60 per task in Artificial Analysis’s runs; Opus 5.5 reaches the same 56 at Xhigh for $3.46, and 58 at Max for $5.98. Pushing Sonnet to its ceiling can cost more than running Opus at a moderate setting for the same result. At Low through High, Sonnet 5.5 is both cheaper and lower-scoring than Opus 5.5 at the same setting, which is the trade the model is built for.
- On Sonnet 5 through the API: the token price is identical, but the switch is not free: run the six-item checklist first, then compare tokens billed per completed task at Low or Medium, where the saving is measured. Output speed in tokens per second is not the same as time to a finished task in an agent loop.
- On Opus 5.5 for everything: move well-scoped tasks, bug fixes, and document work to Sonnet 5.5 at Low or Medium. Keep Opus 5.5 for open-ended work, and for any task you would otherwise run Sonnet at Max. Our Opus 5.5 explainer covers that model.
- In the Claude apps: Sonnet 5.5 runs at Medium by default, the setting where the saving shows. Anthropic has not said which plans get it as their default or whether the Free plan moves from Sonnet 5; check the model picker.
- Security work: expect higher-risk cyber tasks to fall back to Sonnet 5 until the verification program opens.
The bigger picture
Sonnet 5 launched on June 30, 2026, and posted an index score of 38; three months later Sonnet 5.5 posts 56 at the same token price, close behind Opus 5.5’s 58. Our Sonnet 5 guide and Opus 5 vs Sonnet 5 comparison describe the previous generation; our Claude pricing guide will be updated once Haiku 5.5 completes the family. What the effort-level data adds is a rule of thumb for the whole 5.5 family: the tier name tells you the token price, and the effort setting tells you the bill.