Claude Haiku vs Sonnet vs Opus: which model to pick, in the app and on the API

A year ago this question had a simple shape: Sonnet was the everyday model, Opus the expensive one you saved for hard problems. That shape has changed. Anthropic released Fable 5.1, Opus 5.5, Sonnet 5.5 and Haiku 5.5 between 15 September and 7 October 2026, and Opus is no longer the top of the range. Fable 5.1 sits above it, and Anthropic's own docs now say "Most workloads start with Claude Opus 5.5."

So the honest comparison has four rows, not two.

All four have a 1M-token context window, 128K maximum output and a June 2026 knowledge cutoff (Anthropic models overview, checked 22 October 2026). A fifth model, Mythos 5.1, matches Fable's capabilities but is only open to organisations verified through Anthropic's programmes.

Our view: if you pay for Pro, make Opus your default for anything you'll send, publish or decide on, and drop to Sonnet when you want speed. If you build on the API, start with Sonnet and move a task up to Opus only when Sonnet measurably fails it. The two pieces of advice differ because the costs differ: in the app you pay a flat fee, on the API you pay double for every token. Haiku belongs on the API too, for jobs you repeat thousands of times and can test. At a twentieth of Sonnet's price for short prompts, the saving is large enough to be worth testing for.

Open the model picker on the Free plan and you'll find Sonnet and Haiku. No Opus. Pro adds Opus; Max adds Fable, which Pro users can only reach by paying for usage credits (claude.com/pricing, checked 22 October 2026).

Usage on every paid plan works the same way: limits reset on a rolling five-hour window, weekly limits sit on top, and there is "no fixed message count". Anthropic doesn't publish how much faster a long Opus conversation eats into that window than a Sonnet one, so we can't give you a ratio. What we can say is that attachments, long threads and tools all count, and Anthropic warns that Research sessions "can use up your limits faster".

One detail catches people out. Claude's Research feature runs on Opus 5.5 whichever model you've picked, falling back to Opus 4.8 (Claude help centre, Using Research). Choosing Sonnet won't make a research report cheaper on your limits or change the model that writes it.

In practice, on Pro:

• Opus for a contract you need to understand, an analysis someone will act on, a long document where you need every section read properly, or a problem Sonnet already got wrong once.
• Sonnet for drafts, rewrites, quick questions, iterating on code or wording, and anything where you'll send ten follow-ups in five minutes.
• Haiku for a quick rewrite, a one-line formula or a reformatting job, where you'll read the output anyway. On a flat fee its price advantage disappears and what's left is speed, so it rarely earns a place as a default.
• Fable, if you're on Max, for the long jobs Anthropic built it for: research that runs through many steps, or work that should end as a finished spreadsheet or deck.

On Free the choice is made for you, and Sonnet is a strong model to be stuck with.

Whichever model you land on, a more capable one doesn't make checking optional. Count connects to Claude and shows the source, calculation or assumption behind every claim.

Take an illustrative job: 10 million input tokens and 2 million output tokens a month. On Sonnet 5.5 that's $40. On Opus 5.5, $80. On Fable 5.1, $200. On Haiku 5.5, if every prompt stays under 100K tokens, about $2. Batch requests are half price on all of them, and Sonnet's cache reads now cost $0.10 per million tokens (Anthropic models overview).

Does Opus give you twice the quality for twice the price? On the broadest independent measure, not obviously. Artificial Analysis's Intelligence Index, a blend of ten evaluations, puts Opus 5.5 at 58 on its maximum effort setting and Sonnet 5.5 at 56 on the same setting, level with Opus at its "xhigh" setting and above Opus at "high", which scores 54 (Artificial Analysis, checked 22 October 2026). Effort matters almost as much as model choice: a well-configured Sonnet competes with a lightly configured Opus.

Where Opus should pull ahead is in the work Anthropic points it at: multihour coding agents, large refactors, systems engineering, vision-heavy workflows and computer use. Long agent runs are where small per-step differences compound. If your workload is a single call that classifies, drafts or summarises, the index gap is the one to weigh, and it's two points.

So the practical rule: run your own evaluation set through Sonnet first. Move the cases it fails to Opus. Move to Fable only for the long-horizon work Anthropic describes, at two and a half times Opus's price.

Where Haiku pays: the same small job, thousands of times

Take a support queue. Say each ticket is 500 input tokens and the model returns a 20-token label. A million tickets is 500 million input tokens and 20 million output tokens.

• Haiku 5.5: $50 input + $10 output = $60
• Sonnet 5.5: $1,000 input + $200 output = $1,200

Batch processing halves both (Anthropic models overview). The figures are illustrative; the ratio is not. For short prompts Haiku costs a twentieth of Sonnet.

The ratio changes with long prompts. Above 100K tokens Haiku's rate rises to $0.50 / $2.50, so a 200K-token document costs a quarter of what Sonnet charges rather than a twentieth. Still cheaper, much less dramatically so, and long-document reading is exactly the kind of work where the bigger model's extra care pays off.

Haiku 5.5 also moved the floor. Anthropic priced it about 90% below Haiku 4.5 for prompts up to 100K tokens, which puts it at the same per-token price as OpenAI's GPT-6 Luna ($0.10 / $0.50, OpenAI API models page). If you were comparing Anthropic and OpenAI on price for bulk work, that comparison is now level.

Anthropic also lists subagent tasks among Haiku's uses: a larger model plans the work and hands Haiku the many small steps underneath. That split works when the subtasks are narrow and easy to verify, like pulling a field from a document or deciding which tool to call.

A natural assumption: the more capable model is the safer one. On knowledge it's partly true. Opus 5.5 leads Artificial Analysis's AA-Omniscience index at 46 and answers 66% of its 6,000 knowledge questions correctly, against Fable 5.1's 67% (Artificial Analysis, checked 22 October 2026).

The same benchmark also measures something less flattering: when a model doesn't get an answer right, how often it gives a wrong one instead of declining. The lowest rates on that measure belong to small models that mostly decline to answer (MiniCPM5-1B at 1%), not to any Claude tier (Artificial Analysis, checked 22 October 2026). So a more capable model is not automatically a more cautious one: more knowledge doesn't mean more willingness to say "I don't know".

For summarising a document you supply, the newest models aren't on Vectara's leaderboard yet. The previous generation shows how small the differences were: Claude Haiku 4.5 added unsupported facts to 9.8% of summaries, Sonnet 4.6 to 10.6%, Opus 4.5 to 10.9% and Opus 4.7 to 12.0% (Vectara, updated 16 October 2026). The smallest model was no less faithful at that narrow task.

There's no independent accuracy data for Haiku 5.5 yet; anyone quoting a precise figure for it is quoting Anthropic or guessing. Narrow tasks are where Haiku belongs, and the risk there is volume: a 3% error rate is invisible in a demo and is 30,000 wrong labels in a million rows.

A two-minute check before you switch

1. Pick a question you already know the answer to, ideally from your own files, and ask both models.
2. Ask each one: "Which parts of this answer are you least sure of?" The useful answer names specific claims.
3. Ask a question just outside what you gave it. See which model says the documents don't cover it, and which fills the gap.
4. Run the same prompt twice on each. A claim that changes between runs isn't one either model knows.

Before you move a high-volume job to Haiku

1. Label 100 real examples by hand. Include awkward ones: blanks, two categories at once, a field that's missing.
2. Run Haiku and Sonnet on them with the same prompt, and give each an explicit "unknown" option.
3. Read every disagreement. Where the two differ is where the hard cases live.
4. Re-run the 100 next month. Prompts drift, data drifts, and model versions change.

Whichever model you settle on, you'll still be checking its answers. Count connects to Claude through its MCP server and breaks an answer into claims, labelling each as a citation, an assumption, model knowledge or a hallucination, so the check takes minutes rather than an afternoon.

The best argument for Sonnet is speed. Anthropic ranks it "Fast" and Opus "Moderate", and in a working session that difference is the gap between iterating and waiting. For anything conversational, Sonnet usually wins on feel even when Opus would win on a benchmark.

Pick Haiku when:

• The job runs thousands of times and each output is short and checkable.
• Latency matters to a user waiting on the other end, as in a live chat or autocomplete.
• You've measured it against hand-labelled examples and it holds up.
• It's a subtask inside a larger system that something else verifies.

Pick Sonnet when:

• You're on the Free plan, or watching your five-hour window.
• The task is short and you'll check the output anyway: a rewrite, a formula, a first draft.
• You're paying per token and your own tests show Sonnet passes.

Pick Opus when:

• An error would be expensive and hard to spot, as in a financial summary or a policy reading.
• The input is long and every part matters.
• The task runs for many steps, especially coding agents and multi-file changes.
• Sonnet got it wrong and you've already rephrased the prompt once.

The counter-argument to our Opus-by-default advice for Pro users is a fair one: if Sonnet is nearly as good and noticeably faster, defaulting to the slower model wastes time on easy questions. We still think the default should be the model you'd trust with the hard case, because people rarely switch up when a question turns out harder than it looked. They just accept the answer. The same reasoning keeps Haiku off the default: a fast wrong answer to a hard question looks identical to a fast right answer to an easy one.

Is Claude Opus better than Sonnet?

Which Claude model is best for coding?

Can I use Claude Opus on the free plan?

What is Claude Fable, and is it better than Opus?

Which Claude model does Research use?

How much cheaper is Haiku than Sonnet?

Is Claude Haiku good enough to use?

How does Haiku 5.5 compare with GPT-6 Luna?