- Default to Sonnet 5.5. It costs exactly half of Opus 5.5 on input and output ($2 / $10 vs $4 / $20 per million tokens), runs 30%+ faster than Sonnet 5, and lands within a few points of Opus 5.5 on most published benchmarks.
- Escalate to Opus 5.5 for complex, open-ended work that needs sustained judgment: architecture, long ambiguous investigations, anything where a wrong turn is expensive. Anthropic says Opus 5.5 “remains clearly stronger” there.
- Sonnet 5.5 was released on 28 September 2026. Model ID:
claude-sonnet-5-5.
What Changed With Sonnet 5.5
Anthropic released Claude Sonnet 5.5 on 28 September 2026, six days after Opus 5.5 . It is available in the Claude apps, Claude Code, the Claude API, and on Amazon Web Services, Google Cloud and Microsoft Azure. Haiku 5.5 follows in the coming weeks.
The jump over Sonnet 5 is large. Terminal-Bench 4.0 goes from 10.3% to 70.6%, GDPval-AA from 1449 to 1844 Elo, and it is the first Sonnet to finish Pokémon Red working only from screenshots. The price did not move, and it needs far fewer tokens to do the same work, which Anthropic puts at up to 30% less per task.
That combination is what makes the Sonnet vs Opus question worth asking again. With Sonnet 5, Opus 5.5 was the obvious pick for anything serious. With Sonnet 5.5, the gap mostly closed.
Head to Head: Benchmarks
| Evaluation | Sonnet 5.5 | Opus 5.5 | Sonnet 5 | GPT-6 Sol |
|---|---|---|---|---|
| Terminal-Bench 4.0 (agentic coding) | 70.6% | 66.4% | 10.3% | no data |
| FrontierCode 1.1 (agentic coding) | 46.2% (Xhigh: 52.1%) | 54.4% | 42.4% | 49.3% |
| CursorBench 4.0 | 55.5% | 57.8% | 34.1% | no data |
| GDPval-AA v2.1 (knowledge work, Elo) | 1844 | 1846 | 1449 | 1487 |
| AA-Briefcase v1.1 (knowledge work, Elo) | 1811 | 1822 | 1359 | 1483 |
| Humanity's Last Exam (with tools) | 64.5% | 67.7% | 54.9% | no data |
| OSWorld 2.1 (computer use, partial) | 80.1% | 81.8% | 57.0% | no data |
| Chartography (chart reading, no tools) | 61.6% | 64.4% | 15.6% | 53.6% |
Read across the rows and the pattern is clear. Opus 5.5 wins seven of eight, but on most of them by two or three points: 1.7 on OSWorld, 2.3 on CursorBench, 2 Elo on GDPval. Sonnet 5.5 wins Terminal-Bench 4.0 outright. The one real gap is FrontierCode, which checks whether a change could be merged without human edits. There Opus 5.5 leads by 8 points at Max effort, 2.3 at Xhigh.
The FrontierCode footnote matters: Sonnet 5.5 scored lower at Max than at Xhigh because at Max it more often ran Claude Code’s code-review skill across many subagents, which in some tasks caused a timeout or edits outside the task scope. More effort is not automatically better with this model.
Against GPT-6 Sol , which lists at the same $2 / $10, Sonnet 5.5 leads on GDPval-AA, AA-Briefcase and Chartography, and splits FrontierCode depending on effort.
What Each One Costs
| Per million tokens | Sonnet 5.5 | Opus 5.5 |
|---|---|---|
| Input | $2 | $4 |
| Output | $10 | $20 |
| Cache writes | $2.50 | $5 |
| Cache reads | $0.10 | $0.20 |
On the rate card, Sonnet 5.5 is now half price on everything. Until October 7 cache reads were the exception, with both models at $0.20; Anthropic then cut Sonnet 5.5 to $0.10, so long agentic sessions that mostly re-read a cached codebase now get the same 50% saving as output-heavy work.
A worked example: a job with 1M fresh input tokens and 200K output tokens costs $4 on Sonnet 5.5 and $8 on Opus 5.5. In practice the gap is often wider, because Sonnet 5.5 also batches tool calls and tends to finish in fewer steps. Early testers put numbers on that: Lovable saw a third fewer tool calls, Base44 needed 3.6 iterations per app build where Opus 5 took 7.7, and Balyasny measured about 121K tokens per answer against Sonnet 5’s 497K.
To run your own numbers, Sonnet 5.5 is now in our LLM cost calculator and token counter , next to Opus 5.5.
Effort Settings Change the Answer
Both models take an effort setting, and it shifts the comparison. Anthropic’s cost-per-task charts show that on several benchmarks, Sonnet 5.5 at Low or Medium effort beats Sonnet 5’s best score for about a tenth of the cost. On FrontierCode at High effort it scores 10 points above Sonnet 5 at the same setting, at roughly one fifteenth of the cost per task.
Anthropic’s own guidance: Sonnet 5.5 complements Opus 5.5 best at lower effort settings, where it is clearly cheaper per task. At higher settings it can perform comparably to Opus 5.5 at a similar cost, which is the point where you might as well run Opus.
Defaults differ by surface: Medium in Claude Code and the Claude apps, High on the Claude Platform.
A Simple Rule for Picking
Use Sonnet 5.5 for:
- Well-scoped coding: bug fixes, features with a clear spec, test writing, refactors inside one module
- High-volume pipelines: support replies, extraction, classification, code review at scale (CodeRabbit is moving simple and moderate reviews over)
- Documents, slides and spreadsheets; Anthropic highlights its eye for design and template-following decks
- Fast iteration loops where latency matters more than the last few points
Use Opus 5.5 for:
- Architecture and system design, anything open-ended where the model has to decide what to build
- Long, ambiguous investigations that need sustained judgment across many hours
- Tasks where one wrong assumption costs more than the token bill
Or use both. The pattern early testers describe is to let Opus 5.5 set the architecture and let Sonnet 5.5 implement it. In Claude Code that maps naturally onto an Opus planning pass and Sonnet doing the edits.
Migration Notes
- If you run Sonnet with thinking off, switch to the new
between_toolssetting before moving to Sonnet 5.5. It keeps up-front thinking off. - Sonnet 5.5 is the first Sonnet with cyber safeguards: routine bug finding and fixing works, but higher-risk cybersecurity tasks visibly fall back to Sonnet 5.
- Preserved thinking is expanded, so thinking cannot be decoupled from the account that created it. Moving conversations between accounts, including switching accounts mid-session in Claude Code, behaves differently.
- Zero data retention is available, as with Opus 5.5 and Sonnet 5.
FAQ
When was Claude Sonnet 5.5 released?
On 28 September 2026, on the Claude apps, Claude Code, the Claude API, Amazon Web Services, Google Cloud and Microsoft Azure.
How much does Claude Sonnet 5.5 cost?
$2 per million input tokens and $10 per million output tokens, the same as Sonnet 5. Cache reads are $0.10 (since October 7, 2026) and cache writes $2.50. Anthropic says it costs up to 30% less per task than Sonnet 5 because it uses fewer tokens.
Is Sonnet 5.5 better than Opus 5.5?
Not overall. Opus 5.5 leads seven of eight published benchmarks, mostly by two or three points, and Anthropic says it remains clearly stronger at complex, open-ended work. Sonnet 5.5 wins Terminal-Bench 4.0 and costs half as much.
What is the Claude Sonnet 5.5 model ID?
claude-sonnet-5-5 on the Claude API.
Is Sonnet 5.5 better than GPT-6 Sol?
On the benchmarks where both have scores, Sonnet 5.5 leads on GDPval-AA, AA-Briefcase and Chartography. On FrontierCode, GPT-6 Sol’s 49.3% sits between Sonnet 5.5’s Max (46.2%) and Xhigh (52.1%) scores. Both list at $2 input and $10 output.