Claude Haiku vs Sonnet vs Opus: Which Should You Choose?

Which versions does this comparison cover?

Model-family names persist across releases. This guide uses Haiku 5.5, Sonnet 5.5 and Opus 5.5, checked October 8, 2026. If your tool shows an earlier version, read the matching comparison before applying current prices or capabilities.

Sources: Haiku, Sonnet and Opus specifications. Haiku's quoted starting rates apply through 100K input tokens; longer prompts have a higher tier. Equal context capacity does not imply equal reasoning quality, and these capacities are not HaikuChat input limits.

Choose by the shape of the work

A fixed answer from supplied material: extracting a date, classifying a support request or producing a brief summary gives you a clear acceptance test. Start by evaluating Haiku. Require missing values to remain missing, and verify that each output fact can be traced to the input.

Several requirements that interact: planning a project from contradictory notes, interpreting an exception or revising a document under multiple constraints needs more than a fluent summary. Put Sonnet on the shortlist and test whether it preserves every named requirement. The Haiku vs Sonnet comparison examines that trade-off in detail.

A sustained investigation: reconciling several hypotheses, debugging an unfamiliar system or making a recommendation from incomplete evidence merits evaluating Opus. Ask for a traceable recommendation and what evidence would change it. The Haiku vs Opus comparison explains why this needs a different rubric.

These are selection heuristics, not guarantees. A larger model can still fabricate a detail, and a smaller one may be sufficient for a well-specified task. High-consequence outputs need appropriate review regardless of the model name.

The same prompt can have very different economics

At equal usage of 20K input and 2K billable output tokens, the uncached base-rate estimates are $0.003 for Haiku, $0.06 for Sonnet and $0.12 for Opus. The twentyfold and fortyfold gaps apply to this lower-tier, equal-token example. They are not measured job costs or chat subscription prices.

Above 100K input tokens, Haiku's whole-request rates increase to $0.50 input and $2.50 output per million. At 120K input and 2K output tokens, the three estimates become $0.065, $0.26 and $0.52 before caching, discounts or tools. The gap narrows, which is why a blanket “forty times cheaper” claim can mislead a long-document user.

Compare accepted results rather than token rates alone. Keep retries, rejected answers and checking time in the record. If all three satisfy a narrow task equally well, the rate difference is relevant. If only one resolves an essential constraint, the cheaper unsuccessful answers are not equivalent deliverables. HaikuChat pricing separately describes this workspace's credit packs.

Do benchmarks give a complete family ranking?

No single score captures summary fidelity, image extraction, coding and open-ended reasoning. The Haiku 5.5 release evaluation reports Haiku at 39.2% and Sonnet 5.5 at 70.6% on Terminal-Bench 4.0. That is evidence about a specific evaluated setup, not a guaranteed success rate for your task.

That release table does not provide the matching Opus column needed for a clean three-way comparison. We do not fill it using a different version or an unrelated testing setup. The pairwise guides explain the available evidence and its limits.

A model's context window, benchmark score and the tools around it answer different questions. A terminal or computer-use benchmark does not mean HaikuChat runs shell commands or operates a computer. This workspace provides chats, pasted-text summaries and table extraction.

Define when to move to the next tier

Use a written rule: accept a simple extract only if every required value is correct or explicitly missing; escalate a document analysis if it leaves a named conflict unresolved. A rule tied to the task is more useful than deciding after a polished answer arrives.

A workable manual sequence is to try the smallest model likely to pass, inspect the result against the rubric, and use a different tier only for unresolved cases. Track whether escalation actually fixed the problem. If it did not, reconsider the input, instructions or available evidence before repeating the request.

This is a model-selection method, not automatic routing in HaikuChat. To compare another small model, read Haiku vs GPT-6 Luna. To choose among earlier versions, return to the comparison directory. For a short task, open HaikuChat and check the selected model and connection status.

Questions about this comparison