Claude Sonnet and Opus, explained

If your Claude app or your organisation’s Copilot offers a choice between Sonnet and Opus, the decision is narrower than it looks. Both models take in the same amount of your text at once. Opus thinks harder before it answers, and it takes longer doing it.

Where the names come from

Anthropic names its model sizes after forms in poetry and music, ordered by length. A haiku is 17 syllables, a sonnet is 14 lines, and an opus is a whole composition. The three names launched together with Claude 3 in early 2024 and have stuck since.

One common assumption is now out of date. Opus is no longer the largest model Anthropic makes. Claude Fable 5 launched above it on 9 June 2026, and Fable 5.1 replaced it on 1 September at two and a half times the current Opus price on Anthropic’s developer API. The gap at the top has since narrowed: Anthropic says the newest Opus “performs at the level of Claude Fable 5.1 on most work”, which is its claim rather than a measured fact, but it makes Opus the tier most demanding work should try first.

What actually differs

The current pair are Claude Sonnet 5, launched on 30 June 2026, and Claude Opus 5.5, launched on 22 September, which replaced Opus 5 as the current Opus less than two months after Opus 5 came out. Anthropic says Sonnet 5.5 will follow “in the coming weeks”, so expect this pairing to change again before the year is out. Four things separate the current two, and only one of the four is about capacity.

Reasoning depth. Both models work a problem through before replying rather than answering with the first thing that fits. Opus 5.5 does more of that working out, and Anthropic describes it as built “for long-running agentic coding and knowledge work”, agentic meaning work where the model carries out a string of steps on its own. On a knotty multi-step problem you will see the difference; on a request to tidy up an email you will not.

Speed. Anthropic rates Sonnet 5 “fast” and Opus 5.5 “moderate”. The wait is noticeable.

Knowledge cutoff. The date a model’s training data stops, after which it knows nothing unless you tell it or it searches the web. Opus 5.5 is reliable on events to June 2026, Sonnet 5 only to January 2026. That five-month gap is the one reason to pick Opus for an otherwise easy question.

Context window. The amount of text a model holds in mind at once: your whole conversation, anything you have pasted or attached, and the answer being written. The two are identical here, at 1M tokens each.

Tokens, and why that window is smaller than it sounds

Models read tokens rather than words. A token is a chunk of text, roughly three-quarters of a word in English, so a 1M-token window holds less writing than it sounds like. Anthropic puts 1M tokens at about 555,000 words on its current models, against about 750,000 words on those before Claude Opus 4.7, because the newer ones split text into smaller pieces. The number in the specification did not move while the writing it holds fell by roughly a quarter.

The context window is the same on both. Switching to Opus does not let you paste in more. It changes what the model does with what you paste.

Which one to use

Stay on Sonnet and switch to Opus when you can name the reason, such as a long problem that needs sustained reasoning or checking someone’s work closely enough to catch real errors. Anthropic also says Opus 5.5 reads dense charts, diagrams and screenshots “much more precisely” than Opus 5, which matters if your work arrives as images of spreadsheets.

Opus costs you for that. It is slower, which is its own kind of work if you wanted three sentences. Anthropic prices it at twice Sonnet 5 on its developer API, the clearest published measure of the gap, down from two and a half times for Opus 5. On a subscription that reaches you as usage limits rather than a bill, and those moved in your favour in September: Anthropic raised the five-hour usage limits on its Pro, Max, Team and seat-based Enterprise plans when Opus 5.5 launched, and began giving subscribers occasional limit resets. A reset refills your five-hour or weekly allowance when you choose to spend it, from Settings, then Usage, then Reset for free on the web or in Claude Desktop, and it expires on the date shown there.

Whether you get the choice at all is not always yours to make. In Microsoft 365 Copilot the models on offer depend on what your organisation has switched on, and Microsoft adds Anthropic’s models on its own timetable: its roadmap shows Opus 5 arriving in Copilot on 24 July and Fable 5.1 in Cowork and Copilot Studio on 1 September, with no Opus 5.5 listed as of 26 September. See which Copilot is which for what your licence covers.