Claude Sonnet 5.5 arrived on September 28, 2026. If you spend your day asking an agent to fix code, review a pull request, or turn a folder of source material into a useful document, the interesting question is whether it finishes more of that work with fewer detours. The extra 0.5 in its name tells you little about that.
Anthropic's launch post promises output more than 30% faster than Sonnet 5 and up to 30% lower cost on typical tasks. Yet the API price per token is unchanged. Savings have to come from how much work the model does to reach a usable answer. At the highest reasoning setting, Artificial Analysis found the opposite cost result on its evaluation tasks. Both findings can be true; they describe different workloads and settings.
This guide separates Anthropic's claims, independent measurements, and our recommendations. Mengbi did not run its own paid API head-to-head tests for this article. Sources were checked on September 29, 2026.
Specs and API pricing at a glance
| Item | Claude Sonnet 5.5 |
|---|---|
| Release date | September 28, 2026 |
| Claude API model ID | claude-sonnet-5-5 |
| Context / standard maximum output | 1 million / 128,000 tokens |
| Input and output | Text and images in; text out |
| Uncached input / output | $2 / $10 per million tokens |
| Five-minute cache write / cache read | $2.50 / $0.20 per million tokens |
| Default effort | high on the Claude API; medium in Claude apps and Claude Code |
The official model page gives these limits and prices. A separate Batch API beta can generate up to 300,000 output tokens; that is not the standard limit. Anthropic lists the Claude API, Amazon Bedrock, Claude Platform on AWS, Google Cloud, and Microsoft Foundry, subject to each platform's region and account availability.
Sonnet 5 and 5.5 have the same input, output, and cache rates. Opus 5.5 costs twice as much per uncached input and output token. But the price list cannot tell you what one accepted result will cost after retries and review.
Coding improved. How much of that reaches real projects?
In Anthropic's launch benchmark table, Sonnet 5.5 scores 70.6% on Terminal-Bench 4.0, against Sonnet 5's 10.3% and Opus 5.5's 66.4%. On CursorBench 4.0, which uses ambiguous, multi-file tasks from real Cursor sessions, it scores 55.5%, against 34.1% for Sonnet 5 and 57.8% for Opus 5.5. These are results under specified evaluation setups, not a forecast that seven in ten bugs in your repository will be fixed.
The useful question for a team is whether the agent reads the right files, makes a scoped change, runs a check, and repairs a failure before asking for review. Anthropic itself says Opus 5.5 remains stronger on difficult, open-ended work that requires sustained judgment. More reasoning does not automatically help: on FrontierCode 1.1 Main, the official table gives Sonnet 5.5 52.1% at xhigh but 46.2% at max. The launch footnote says extra review steps led to timeouts or edits beyond scope in some cases.
A closer look at pull-request review
CodeRabbit's September 28 test used the same 13 known-bug cases for Sonnet 5.5 and Sonnet 5 inside its review pipeline. With thinking on, the new model caught 6 of 13, compared with 4 of 13 for Sonnet 5, at 41.2% versus 40.0% actionable precision. It also missed two issues that Sonnet 5 caught. On the same hard cases, Opus 5.5 caught 8 at Standard and 10 at Max.
On a wider set of 44 open-source pull requests, Sonnet 5.5 averaged 6:33 per review versus 13:31 for Sonnet 5 and produced fewer comments. Scoring of known issues in that 44-PR set was still pending when CodeRabbit published, so it supports a speed and volume observation, not a claim that the model caught more bugs there. The results also belong to CodeRabbit's particular review pipeline; they are not a universal benchmark for coding agents.
Same token price, two very different cost stories
Anthropic says Sonnet 5.5 typically uses fewer tokens to finish work and can cost up to 30% less per task than Sonnet 5 in its tests. Artificial Analysis tested all five effort levels with the provider's default fallback enabled. At max, Sonnet 5.5 reached 56 on its Intelligence Index, up 18 points from Sonnet 5 and just below Opus 5.5's 58. But it generated roughly 193,000 output tokens per evaluation task and cost $7.60 per task under that organization's accounting, about 50% more than Sonnet 5. That does not mean an ordinary chat message costs $7.60; it describes a demanding benchmark mix at the highest effort setting.
Artificial Analysis says fallback occurred in around 0.1% of its tasks, mostly on Terminal-Bench. It also notes that its prerelease deployment had a structured-output issue that Anthropic later fixed; some evaluations may be rerun. So treat 56 as a measured score for a named configuration and version, not a promise for the default settings in every product.
For an honest cost comparison, start medium on well-specified fixes, try high on harder cross-file work, and reserve xhigh or max for cases where more deliberation earns its price. Record acceptance rate, wall-clock time, input and output tokens, cache usage, retries, and human correction time. A cheaper first response that needs three repair rounds can lose to a more expensive one that ships; the reverse happens too.
Documents and charts need their own acceptance test
Anthropic reports 1844 Elo for Sonnet 5.5 on GDPval-AA v2.1, close to Opus 5.5's 1846 and above Sonnet 5's 1449. On AA-Briefcase v1.1, the official scores move from 1359 to 1811. Elo is a relative score, not a percentage of correct answers. The launch post also shows document, slide, spreadsheet, and visual-understanding examples, selected by the vendor.
To test this work yourself, give the old and new models the same source documents and a concrete checklist. For a board deck, check every requested slide, trace each number back to its source, inspect chart axes, and count the edits needed before delivery. For financial, legal, or other high-stakes material, keep appropriate human review. Artificial Analysis also found that Sonnet 5.5 still trails Opus 5.5 on its factual-knowledge and scientific-reasoning tests.
API migration: the model ID is only the first change
After switching to claude-sonnet-5-5, inspect the official migration notes for your integration:
- Turning off up-front thinking changed.
thinking: {"type": "disabled"}now returns a 400. Usebetween_toolsatlow,medium, orhigh;xhighandmaxrequire adaptive thinking. - Forced tool choice is unsupported.
tool_choicetypesanyand a namedtoolreturn 400. Useautoand, where appropriate, strict tool schemas or structured outputs. - Stored thinking blocks are bound to a model and conversation. Replaying them after changing earlier messages, tools, or system instructions can fail on affected accounts. Review your history-editing flow.
- Older computer-use tool versions need attention. The Claude API and Google Cloud reject
computer_20251124; move tocomputer_toolset_20260801. Amazon Bedrock has different compatibility. - Progress text between tool calls may move into
thinkingblocks. A UI that only renderstextcan appear to stall unless it handles the display setting correctly.
Effort levels have also been recalibrated, so do not assume Sonnet 5's high spends the same time or tokens on Sonnet 5.5. Some high-risk cybersecurity requests can be refused or fall back; an HTTP 200 alone does not prove a task finished. Check stop_reason and the response details.
Should you choose Sonnet 5.5 or Opus 5.5?
Start with Sonnet 5.5 for frequent tasks with a clear finish line: routine bug fixes, a first review pass, document extraction, and template-driven writing. Official and third-party data both make a credible case for better capability and speed. Use your own acceptance set before changing all production traffic.
Keep Opus 5.5 in the comparison for open-ended, costly-to-miss work. CodeRabbit's hard review cases and Anthropic's own positioning show that the flagship still has a role. Routing tasks by difficulty and measuring cost per accepted result will tell you more than choosing a winner from one leaderboard row.
You can start with the Claude tool profile, read our Opus 5.5 analysis, or compare AI coding agents for the surrounding workflow.
Frequently Asked Questions
When was Claude Sonnet 5.5 released?
Anthropic dates the release September 28, 2026. Sonnet 5 launched in June 2026. This article and its cover concern 5.5.
How much does the Sonnet 5.5 API cost?
Claude API rates are $2 per million uncached input tokens, $10 per million output tokens, and $0.20 per million cache-read tokens. These are API rates, not Claude subscription prices; cache writes, batch processing, tools, and other platforms have their own terms.
Is it 30% cheaper than Sonnet 5?
The list rates are unchanged. Anthropic's “up to 30%” saving is a task-level result under its tests. Artificial Analysis measured a higher task cost at max on its own benchmark. Your result depends on workload and effort.
Can Sonnet 5.5 generate images?
Its official input and output specification is text and images in, text out. It can understand an image and write code or a description, but it is not a native image-generation model.
Sources and scope
- Anthropic launch announcement: release date, vendor benchmarks, speed and cost claims, positioning, and footnotes.
- Anthropic model specifications and migration notes: API ID, prices, limits, platforms, and integration changes.
- Artificial Analysis evaluation: independent index, effort-level costs, token use, and prerelease limitation.
- CodeRabbit review test: known-bug cases, pull-request review timing, and methodology limits.
Last verified September 29, 2026. The cover was supplied by the reader; it is not presented as an Anthropic-licensed asset or evidence of model performance. Prices, rankings, and platform support may change; consult the linked sources for current details.