Claude Sonnet 5.5, which Anthropic released on September 28, costs half as much as the flagship Opus 5.5 and barely trails it on benchmarks: on "office" knowledge work the gap is 2 Elo points out of 1800+.
So which is better, Opus 5.5 or Sonnet 5.5? The short answer: Opus is worth the money when the task is vague and long, and Sonnet is enough for everything else. Below are prices, Anthropic's test results and a short rule for choosing.
What they have in common
Opus 5.5 came out on September 22, Sonnet 5.5 six days later. On specs the models are twins: a 1M-token context window (about 555,000 words), up to 128,000 tokens of output, text and image input, and a knowledge cutoff of June 2026 for both.
They differ in price, speed and how they "think":
| Parameter | Sonnet 5.5 | Opus 5.5 |
|---|---|---|
| Speed | fast | moderate |
| Reasoning | adaptive, can be reduced | adaptive, always on |
| Default effort level in the API | high | medium |
| Price on GPTunneL, per 1K input / output | $0.004 / $0.02 | $0.012 / $0.06 |
Look at the last row: on our price list Opus costs three times as much as Sonnet. Current prices are on the pricing page.
What the tests showed
Anthropic compared the two models on eight benchmarks, and the Opus 5.5 announcement also includes results for the previous flagship, Opus 5. Together they show where Sonnet 5.5 stands relative to its two neighbors:
| Test | What it measures | Sonnet 5.5 | Opus 5.5 | Opus 5 |
|---|---|---|---|---|
| Terminal-Bench 4.0 | agent work in the command line | 70.6% | 66.4% | 52.3% |
| FrontierCode 1.1 | real programmers' tasks | 52.1% | 54.4% | 48.0% |
| CursorBench 4.0 | edits across multiple files | 55.5% | 57.8% | 46.6% |
| GDPval-AA | professional tasks, Elo | 1844 | 1846 | 1708 |
| AA-Briefcase | long document tasks, Elo | 1811 | 1822 | n/a |
| Humanity's Last Exam | hard questions with tools | 64.5% | 67.7% | 63.6% |
| OSWorld 2.1 | computer use | 80.1% | 81.8% | 74.0% |
| Chartography | reading charts | 61.6% | 64.4% | n/a |
The Opus 5 figures come from the Opus 5.5 announcement. "n/a" means the announcement has no comparable measurement: it does not report AA-Briefcase, and Opus 5's Chartography score was measured with tools, while the other two models were measured without them.
The main takeaway: Sonnet 5.5 beats the previous flagship Opus 5 on all six tests where data exists, even though on our price list it costs $0.004 and $0.02 per 1K tokens versus $0.015 and $0.075 for Opus 5. The new Opus 5.5 stays ahead by a few more points, but by a small margin over Sonnet.
Why Sonnet leads on Terminal-Bench
This is the only row where Sonnet beats Opus 5.5, by 4.2 points. There are several possible reasons, and not all of them are proven:
- Noise. Anthropic gives the standard error of Opus 5.5's score on this test as ±2.6 points. The gap is less than two such errors, so it can't be called reliable.
- Safeguards. The refusal hypothesis is partly supported. Opus 5.5 was tested with production safeguards on, and a general footnote from Anthropic says that when the safeguards kicked in, cyber tasks were completed by the older Opus 4.8 and biology tasks by Opus 5. According to Anthropic, this likely lowers Opus 5.5's scores. Terminal-Bench isn't mentioned separately, and the footnotes for Sonnet's tests have no such caveat, so the safeguards' contribution to the gap remains a guess.
- Different settings. Opus was run at the xhigh effort level. The level for Sonnet on this test is not stated.
The table as a whole shouldn't be read as a fair duel. On FrontierCode, Sonnet's 52.1% was reached at xhigh, while at the maximum level, max, it scores only 46.2%: Anthropic explains that at max the model more often spawned subtasks, and in two cases they hit timeouts. In addition, two of the Elo ratings were measured by Artificial Analysis on a pre-release version with a bug that could have understated Sonnet's score. The bug has since been fixed.
Anthropic itself puts the conclusion this way: Opus 5.5 remains clearly stronger at complex, open-ended work that requires sustained judgment. That matches the table: the biggest gap is on Humanity's Last Exam, which tests reasoning rather than execution.
Where Sonnet wins
Speed. Anthropic classes Sonnet 5.5 as "fast" and Opus 5.5 as "moderate" in latency. Wiegold's measurements back this up: Sonnet wrote the chess engine in 5 minutes against 7 for Opus, and the landing page in 8 minutes against 9 (one run per model, so the numbers are indicative). For a chat where you wait for the answer, and for tasks that run in batches, the difference is noticeable.
Cost per task. Take a typical request: 100 thousand input tokens and 10 thousand output tokens, a big document and a detailed answer. The calculation assumes an equal number of tokens; in practice the models use them differently, and real savings can be smaller (see below).
| Model | On GPTunneL |
|---|---|
| Sonnet 5.5 | $0.60 |
| Opus 5.5 | $1.80 |
Across a hundred such requests, the gap is $120. If Sonnet gets the job done, there is no reason to overpay for Opus.
Which model to pick for your task
Anthropic's numbers are its own measurements. In the two weeks since release, third-party tests have appeared, and they diverge by domain: in some areas Sonnet barely trails, in others the gap is visible.
Code and review
This is where Sonnet 5.5 is closest to Opus. According to Anthropic, it is even ahead on Terminal-Bench (70.6% vs 66.4%, though the effort levels differ), while on FrontierCode and CursorBench it trails by 2-3 points.
The code review service CodeRabbit tested the models on 13 hard bugs from real open-source projects. Sonnet 5.5 found 6, Opus 5.5 in standard mode found 8, and at maximum effort 10. The authors' verdict: "the most convincing Sonnet we have tested", but not a replacement for Opus: four hard cases defeated every Sonnet configuration. The sample is small, and the authors themselves advise checking on your own code.
Developer Thomas Wiegold asked both models to write a chess engine in Go. Both engines passed the independent check, but two of Sonnet's own tests failed: it had invented the expected values. In the Hacker News discussion, users broadly agree that in agentic coding Sonnet is almost level with Opus.
Takeaway: routine edits, bugs with clear tests and ordinary review - Sonnet. Hard bugs and code that runs unattended - Opus.
Graphics, design and interfaces
Wiegold asked both models to build a one-page site for a beach film night. Opus took first place (9 animations, the most accessibility work), while Sonnet stayed "close behind" (4 animations, its own print-poster style).
The MindStudio team tested graphics on games and a landing page. In a 3D game Opus wasn't clearly better than Sonnet overall; it only won on the character creator. In a Breaking Bad-style game Opus had the best-developed faces, and on a watch-launch landing page the most polished scroll animation.
Takeaway: an ordinary page or interface is fine with Sonnet. Opus pays off in finish: animation, faces and characters, accessibility. Both tests are single runs and subjective.
Finance, documents and analytics
MindStudio asked the three models to find 8 fraud indicators planted in the data. All three found all of them, but Opus cost 4-5 times more, and its report was better formatted, not better in substance. On the GDPval-AA office-work ranking, the gap between Sonnet and Opus is 2 points (1844 and 1846).
Takeaway: spreadsheets, documents, data reconciliation, analytics - Sonnet. Take Opus if how the report looks matters.
Hard and long tasks
Here the difference is real. Anthropic itself says Opus 5.5 remains clearly stronger at complex, open-ended work that requires sustained judgment. The biggest gap in the table above is Humanity's Last Exam (67.7% vs 64.5%). CodeRabbit confirms this on the hardest bugs. Wiegold adds a practical detail: the values Sonnet invented in its tests are dangerous where an agent works unattended, and Opus's extra care was worth the money.
Takeaway: architecture, research, audits and agents that run for hours - Opus.
Security and low-level code
The most common complaint in the Hacker News discussion is not about quality but about the "cyber" safeguards. Users write that they trigger on ordinary code (debugging a parser, checking a write-ahead log) and the conversation falls back to an older model. These are not nitpicks: the announcement says higher-risk security tasks fall back to Sonnet 5 with Sonnet 5.5.
Takeaway: if you work on security or low-level code, check the behavior on your own task before betting on the model.
Sonnet 5.5 vs Opus 5.5: what to pick and where
| Task | What to pick |
|---|---|
| Routine code, edits, ordinary review | Sonnet 5.5 |
| Spreadsheets, documents, analytics | Sonnet 5.5 |
| Landing pages and interfaces without special requirements | Sonnet 5.5 |
| Animation, characters, polish | Opus 5.5 |
| Hard bugs, unattended code | Opus 5.5 |
| Architecture, research, audits, long-running agents | Opus 5.5 |
| Security, low-level code | check on your own task |
One more practical argument for Sonnet: a Hacker News user writes that they hit session limits on Opus and see Sonnet 5.5 as a way to stretch them.
How to use both
The most practical scheme is to pick the model as you go. Start with Sonnet: it is faster and cheaper. If it misses twice or the task turns out harder, switch to Opus. One of Anthropic's partners, developer Creator, described the reverse order too: they would trust Sonnet 5.5 to implement a game framework that Opus 5.5 designed. That makes a division of roles: Opus thinks, Sonnet does.
If you switch models automatically in code, put a cap on the number of attempts, or you will get a loop where the cheap model keeps handing the task to the expensive one.
Where to try them
Both models are on GPTunneL: Claude Sonnet 5.5 and Claude Opus 5.5. In the API they are called claude-5.5-sonnet and claude-opus-5.5. No VPN needed, local payment methods, no subscription, one shared balance. The easiest way to compare is on your own task in one chat: send the same request to both and look at the answer and the cost.
Sources
- Claude Sonnet 5.5 announcement, Anthropic, 09/28/2026 - prices, benchmarks, customer feedback.
- Claude Opus 5.5 announcement, Anthropic, 09/22/2026 - Opus 5.5 and Opus 5 results, footnotes on safeguards and error margins.
- Claude Sonnet 5.5: documentation overview, Anthropic - specs, prices, comparison with Opus 5.5.
- Sonnet 5.5 for code review, CodeRabbit, 09/29/2026 - a test on 13 known bugs.
- Claude Sonnet 5.5 Review: I Tested It Against Opus 5.5, Thomas Wiegold, 09/30/2026 - a comparison on design and code.
- Claude Haiku 5.5 vs Sonnet and Opus 5.5, MindStudio, 10/09/2026 - a comparison on practical tasks.
- Sonnet 5.5 discussion on Hacker News, 09/28/2026 - user feedback.