Claude Opus 5.5 vs Sonnet 5.5 vs Haiku 5.5: which to use

Use Claude Sonnet 5.5 by default, Claude Haiku 5.5 for high-volume simple work, and Claude Opus 5.5 only where your own tests show Sonnet falls short. Opus costs $4 and $20 per million input and output tokens, Sonnet $2 and $10, Haiku $0.10 and $0.50, and on Epoch AI's capability index Opus and Sonnet are within each other's range of doubt.

All 3 have the same 1 million token context window, 128,000 output tokens and June 2026 knowledge. What differs is price, speed, a few controls, and how they behave on long, hard tasks.

Side by side

Scores are from the Epoch Capabilities Index by Epoch AI (CC BY 4.0), which combines over 50 benchmarks. Opus's 90% range runs from 163.5 to 171.0 and Sonnet's from 161.4 to 168.3, so a general benchmark cannot tell them apart. Haiku's 159.0 is lower but well above GPT-5's 150, and above Claude Sonnet 4.5's 146.8. Each model's page has the details: Opus 5.5, Sonnet 5.5, Haiku 5.5.

What each is for

Sonnet 5.5 is the everyday model: code generation, data analysis, writing, tool-using agents. Its default effort is high, so lowering effort is the first way to make it faster and cheaper before you switch models.

Haiku 5.5 is for the many small calls: labelling tickets, pulling fields out of documents, choosing which tool or model handles a request, summarising, and sub-agent tasks inside a larger agent. Anthropic suggests starting here and upgrading only where tests show a gap. Two cautions: requests over 100,000 input tokens cost 5 times as much, and the same text counts about 30% more tokens than on Haiku 4.5, so recount before comparing bills.

Opus 5.5 is for work that runs for hours: autonomous coding across a large codebase, big refactors, complex systems work. On a short question it is usually slower and dearer for the same answer.

Differences that break code

The 3 models do not take the same requests:

• Thinking. Opus 5.5 always thinks; a request with thinking: {"type": "disabled"} returns a 400 error. Sonnet 5.5 also rejects disabled and offers between_tools as its lowest setting. Haiku 5.5 still accepts disabled. On all 3, set depth with effort, not a token budget.
• Forced tool use. tool_choice of any or tool returns a 400 error on Opus 5.5 and Sonnet 5.5. Use auto with strict tool schemas instead.
• Sampling. temperature, top_p and top_k are rejected, as our article on their removal explains.

So "just swap the model ID" is not always safe. Run your eval set on each candidate.

A worked month

A product search assistant: 50,000 requests a day (1.5 million a month), 3,000 input tokens and 500 output tokens each.

• Opus 5.5: 1.5 million × (3,000 × $4 + 500 × $20) ÷ 1 million = $33,000
• Sonnet 5.5: 1.5 million × (3,000 × $2 + 500 × $10) ÷ 1 million = $16,500
• Haiku 5.5: 1.5 million × (3,000 × $0.10 + 500 × $0.50) ÷ 1 million = $825

Now route: a Haiku call (or a simple rule) decides whether a request is easy. If tests show Haiku handles 70% well, send those to Haiku and the rest to Sonnet: 0.7 × $825 + 0.3 × $16,500 = $5,528, a third of all-Sonnet. Thinking tokens are billed as output, so lower effort on the easy share saves more again.

Try itOpus, Sonnet or Haiku: what a month costs

Requests a day

50,000

Input tokens a request

3,000

Output tokens a request

500

Requests Haiku can take

70%

Right awayBatch (50% off)

Claude Opus 5.5 score 166.8$33,000

Claude Sonnet 5.5 score 164.5$16,500

Claude Haiku 5.5 score 159$825

70% Haiku, 30% Sonnet $5,528

Routing saves $10,973 a month against all on Sonnet

Combine them

Anthropic describes 2 patterns that put most tokens on a cheaper model:

• Orchestrator and workers. A strong model plans and checks; Haiku sub-agentsSubagentAn agent started by another agent for one part of a task. It works in its own fresh context and returns only its findings. do the bulk reading and extraction.
• Executor and advisor. A cheaper model does the work and asks a stronger one only at hard decisions.

Either way, the bill follows the model that reads the most tokens, so put the cheap model there. Prompt caching cuts the rest: cache reads are 5% of the input price on Opus and Sonnet. The prompt caching calculator prices your prompt on each model, and which model finds the cheapest model that scores as well as you need.

What to do

1. Start new work on Sonnet 5.5; try lower effort before a smaller model.
2. Move high-volume, well-defined steps to Haiku 5.5 once tests pass.
3. Reserve Opus 5.5 for long agentic jobs where Sonnet fails your tests.
4. Re-run the comparison at each release: the gaps between tiers keep shrinking.