Model Selection: When to Use Sonnet, When to Use Opus

The issue: Most researchers use whichever model Claude defaults to and never think about it. That is often fine — but sometimes wrong. Different Claude models differ meaningfully in capability, speed, and cost. Knowing when those differences matter, and when they do not, is worth understanding once.

The three models

Claude currently has three models in active use:

Sonnet is the default in both Claude Desktop and Claude Code. It is the right choice for most tasks most of the time. The question is when to upgrade to Opus — or, occasionally, when to downgrade to Haiku.

Where you can actually choose

Claude Desktop (claude.ai): Pro subscribers can switch models using the model selector in the interface, visible before starting a conversation. The default is Sonnet. You can switch to Opus for a single conversation without affecting the default.

Claude Code: The default model is Sonnet. You can override it:

• In a single session: claude --model claude-opus-4-8

• In settings: add "model": "claude-opus-4-8" to your .claude/settings.json

Haiku in Claude Code: claude --model claude-haiku-4-5-20251001

When the difference matters

Use Opus when:

The task requires genuine reasoning about difficult material

Reading a dense theoretical text, evaluating a complex argument, making a judgment call that requires holding many considerations in tension — these are tasks where Opus noticeably outperforms Sonnet. Not always, but often enough that it is worth switching.

Examples:

• "Evaluate whether this argument about X is consistent with what the source says in section 3 and what we established in note Y"

• "Is this a genuine counterexample to my thesis, or does it fit within my framework if reframed as..."

• Long-range coherence in a document — tracking whether Chapter 5 is consistent with the claims made in Chapter 2

The task involves subtle language judgment

Editing academic prose for register and precision, not just grammar. Evaluating whether a translation preserves the rhetorical structure of the original. Judging whether a claim is overstated.

You are doing high-stakes interpretation and want a second, better reader

If Sonnet's reading of a source does not feel quite right — if it seems to be missing something or settling for the obvious — switch to Opus and ask again. Opus is not always right where Sonnet is wrong, but its interpretive range is wider.

The task involves extended structured output that must hold together internally

Generating a long, multi-part document where later sections must be consistent with earlier ones. Opus maintains coherence over longer outputs better than Sonnet.

Use Sonnet when (i.e., most of the time):

• Extraction tasks: finding names, dates, places, events in a document

• Conversion and formatting: markdown, DOCX, citations

• Summarisation of clearly-written texts

• Writing assistance where the content is yours and Claude is editing

• Answering specific factual questions about a document

• Batch processing: running the same operation over many sources

• Anything that is mechanical or pattern-based rather than judgment-based

Sonnet on these tasks is not "worse" — it is appropriate. The difference between Sonnet and Opus on a table extraction task is negligible. The cost and speed difference is not.

Use Haiku when:

Haiku is rarely the right choice for research tasks in practice, because the quality gap on anything non-trivial is significant. The main cases where it makes sense:

• Simple, very high-volume tasks where you are paying per token (API, not subscription)

• Quick classification: "does this document mention X or not?" over a large corpus

• Tasks where the output is entirely determined by clear rules with no judgment involved

On a subscription plan (Pro, Max), you are not paying per token — you are spending quota. Haiku's lower quota cost is only relevant if you are regularly hitting rate limits on Sonnet. Most researchers do not hit limits this way.

The quota angle

On Pro and Max subscriptions, heavier models consume your quota faster. Running a long Opus session on a large corpus can exhaust your hourly quota in ways a Sonnet session on the same corpus would not.

Practical implication: Use Opus surgically — for the specific hard interpretive task — rather than as your session default when doing batch processing. Switch back to Sonnet for extraction and formatting even within the same project.

If you regularly exhaust your quota doing intensive work, see A.token.management for strategies. Model selection is one lever; session structure and document preparation are others.

A decision heuristic

Ask yourself: is this task hard?

Specifically: Would a capable, knowledgeable research assistant find this task intellectually demanding — or would it be routine?

• Routine (format, extract, convert, summarise clear text) → Sonnet

• Demanding (reason, judge, evaluate, interpret, hold together) → Opus

A second heuristic: If Sonnet's output leaves you unsatisfied and you can't diagnose why, switch to Opus before assuming the problem is your prompt. Sometimes the model is the limiting factor.

What does not change between models

• Your data still goes to Anthropic's servers regardless of which model you use. Model selection does not affect data privacy. See A.issue.privacy-security.

• Context window: Sonnet 4.6 and Opus (4.6 and later) have 1,000,000 tokens; Haiku 4.5 has 200,000 tokens. For most research tasks the difference does not matter — 200,000 tokens holds several full-length monographs.

• The tools available (file access, web search, etc.) are the same.

Quick reference

Related

• A.token.management — managing quota and rate limits; model selection is one lever among several

• A1.desktop-vs-code — where you can and cannot select models depending on which tool you use

• A4.conceptual-vocabulary — what tokens and context windows are; background for understanding why models differ in cost

• A.issue.context-scope — often the right response to poor Sonnet output is better context scope, not a model upgrade; check this first

• A.issue.costs-quotas — the economic dimension of model choice: Pro quota, API pricing, when each model draws more heavily on your allowance