Which Claude Model Should Lawyers Use? Opus vs Sonnet vs Haiku

The Short Answer

Each Claude model fits a different type of legal work. Choosing wrong costs you either quality or speed, sometimes both. A 2025 Thomson Reuters Institute survey found that 78% of law firm respondents believe generative AI will become central to their workflow within five years, yet just 40% said their organizations provide any form of gen AI training to staff. Matching each model to its task is the fastest way to close that gap.

Here is the breakdown:

• Opus 5.5: Deep legal reasoning. Contract analysis, multi-step research, complex memo drafting, ethical analysis. Default to this for substantive work.
• Sonnet 5.5: Fast and competent. Deposition summaries, first-draft motions, research outlines, client communication drafts.
• Haiku 4.5: Speed and volume. Email drafts, formatting, data extraction, billing narratives, simple Q&A.

The rest of this guide explains when each model earns its place, with use cases, cost considerations, and a decision framework you can apply today.

Understanding Claude's Model Tiers

Opus, Sonnet, and Haiku are tiers of the same Claude model family. The practical differences are capability, speed, and cost. As of October 2026, Anthropic's current lineup is Claude Fable 5.1 (the flagship, released September 1), Opus 5.5 (released September 22), Sonnet 5.5 (released September 28), and Haiku 4.5. Opus 5 and Sonnet 5 are still available as legacy models. The Opus guidance below applies equally to the flagship tier. Anthropic lists Haiku 4.5's retirement as not sooner than October 15, 2026 and has said Claude Haiku 5.5 will follow in the coming weeks; the Haiku guidance below applies to whichever Haiku model is current.

Think of it like hiring: Opus is the senior partner who catches what everyone else misses. Sonnet is the experienced associate who drafts clean work product fast. Haiku is the sharp paralegal who processes volume without complaint.

All three models are available on Claude's paid plans, including Team and Enterprise, the plans we recommend for client work. Team and Enterprise are commercial products, and Anthropic does not use inputs or outputs from its commercial products to train its models by default, regardless of which model you select. Enterprise also adds custom data retention controls, which Anthropic's pricing page does not list for Team. For a full cost comparison, see Claude AI Pricing for Lawyers.

Opus 5.5: The Powerhouse for Legal Reasoning

Opus is the most capable of the three tiers covered here (Anthropic's Fable tier sits above it). It handles the work that demands precision, nuance, and multi-step analysis: the work where mistakes have consequences.

Opus excels at:

• Complex contract review: Analyzing a 50-page M&A agreement for non-standard indemnification provisions, identifying conflicts between representations and disclosure schedules, and drafting redline language. See our full walkthrough in AI Contract Review with Claude.
• Multi-factor legal tests: Applying a 7-factor balancing test to a trade secret misappropriation claim, weighing each factor against the facts, and producing a structured analysis memo.
• Research memos: Synthesizing conflicting authority across jurisdictions and identifying the strongest arguments on both sides.
• Ethical analysis: Navigating conflicts of interest, privilege questions, and regulatory compliance issues where the answer depends on multiple overlapping rules.

Research from Stanford CodeX points the same way. In a December 2024 CodeX experiment on encoding insurance policies and adjudicating claims, encodings from OpenAI's o1-preview reasoning model produced correct adjudications about 83% of the time on average, compared with about 27% for GPT-4o. The study tested OpenAI models, not Claude, but it shows how much the reasoning step can matter on multi-step legal tasks.

When the stakes are high or the analysis requires holding multiple threads at once, Opus pays for itself. A partner billing at $800/hour who saves three hours on a contract review has already covered a month of Claude Team subscription costs.

Sonnet 5.5: Fast and Capable for Everyday Tasks

Anthropic positions Sonnet as the best combination of speed and intelligence in its lineup, with lower latency and half the per-token price of Opus. For tasks that need competence but not Opus-level depth, Sonnet is the right choice.

Sonnet handles:

• Deposition summaries: Summarizing a 150-page deposition transcript into organized sections with key admissions flagged. See our guide on Claude for Litigation and Deposition Summaries.
• First-draft motions: Drafting a standard motion to compel or motion to dismiss that you refine and finalize.
• Research outlines: Mapping out the legal landscape for an issue before diving deep with Opus.
• Client communications: Drafting status updates, engagement letters, and case summaries in plain English.
• Discovery review: Sorting and categorizing documents by relevance, privilege, and issue tags.

Sonnet is a strong everyday workhorse. Its speed makes it well suited to iterative drafting: generate a first draft, revise the prompt, and generate again without a long wait. Sonnet typically returns a solid draft noticeably faster than Opus. For tasks where you plan to edit the output anyway, that speed advantage compounds across dozens of daily tasks. For prompt techniques that work across all models, see the CRAFT Framework.

The biggest productivity gains come not from using the most powerful model available, but from using the right model for each task. For routine drafting and summarization, Sonnet is often all you need, which leaves Opus for the work where depth matters.

Haiku 4.5: Speed and Volume

Haiku is the fastest Claude model and the lowest cost per token. It processes simple tasks at high volume without delay.

Haiku works best for:

• Email drafting: Generating 20 routine client status update emails from a template and case notes.
• Document formatting: Converting documents between formats, applying consistent styling, restructuring sections.
• Data extraction: Pulling key terms from structured forms such as intake questionnaires, insurance declarations, closing checklists.
• Billing narratives: Cleaning up time entries into professional billing descriptions.
• Simple Q&A: Answering straightforward factual questions where deep analysis is unnecessary.

A solo practitioner processing 30 new client intake forms each week can run them all through Haiku quickly, extracting names, contact details, case types, and conflict-check data into a structured spreadsheet. Done manually, the same task can take hours.

Anthropic describes Haiku 4.5 as the fastest model with near-frontier intelligence, but for complex legal analysis we recommend Opus or Sonnet, the more capable tiers. Use Haiku for tasks where speed and cost matter more than analytical depth.

On Anthropic's published API pricing, Haiku 4.5 costs $1 per million input tokens and $5 per million output tokens, against $4 and $20 for Opus 5.5 and $10 and $50 for Fable 5.1. Routing routine administrative tasks to Haiku instead of Opus cuts per-token API costs by about three quarters.

Comparison Table: Opus vs Sonnet vs Haiku for Legal Work

How to Choose the Right Model for Each Task

Use this decision framework when selecting a model:

• Will this go to a court or opposing counsel? → Opus. The stakes justify the thoroughness.
• Is this a first draft you will heavily edit? → Sonnet. Speed matters more than perfection on first passes.
• Is this routine, repetitive, and low-stakes? → Haiku. Save your budget for work that demands it.
• Are you analyzing a complex legal question with multiple factors? → Opus. Multi-step reasoning is its strength.
• Are you summarizing or extracting information? → Sonnet. It handles synthesis well at faster speeds.

Default to Opus for substantive work; switch to Sonnet or Haiku for speed or volume. Opus costs more (at API rates, twice Sonnet's and four times Haiku's per-token price), but in our view that premium is usually worth paying when complex analysis needs to be right the first time.

If you are unsure, start with Opus. Anthropic's help center says you can change the model at any point in a conversation, so you can move to a faster model if the task turns out to be simpler than expected. The reverse, discovering that Haiku missed a critical issue, is far more costly than starting with the slower model.

For prompt strategies that maximize output quality on any model, see our 50 Claude AI Prompts for Lawyers and the Legal Prompt Engineering Guide.

A Practical Workflow Using All Three Models

The most efficient approach uses all three models in a single matter. Here is how a tiered workflow looks in an M&A due diligence context:

1. Haiku, data extraction: Process 50 ancillary documents (employment agreements, vendor contracts, lease agreements). Extract key terms: parties, dates, termination provisions, change-of-control clauses, assignment restrictions. Haiku handles this quickly at low cost.
2. Sonnet, summarization: Take each extracted dataset and produce a one-page brief per document. Flag unusual terms. Generate a master index of all agreements with key dates and obligations.
3. Opus, analysis and risk assessment: Analyze the full acquisition agreement against the extracted data. Identify conflicts between representations and the ancillary documents. Draft a risk matrix. Produce redline recommendations for the purchase agreement. Draft the due diligence memo.

This tiered approach maximizes quality where it matters and speed where it does not. The Haiku and Sonnet steps run quickly, which leaves Opus for the substantive analysis that would otherwise take the most associate time.

The same pattern applies across practice areas. In litigation, use Haiku to extract dates and parties from discovery productions, Sonnet to draft privilege logs, and Opus to analyze key documents for case strategy. In real estate, Haiku pulls lease terms from a portfolio of agreements, Sonnet summarizes each lease, and Opus identifies the provisions that need renegotiation before closing. The principle holds: delegate volume to Haiku, competence to Sonnet, and judgment to Opus.