Claude Opus 4 vs Sonnet 4 for Immigration Attorneys: Task Guide

Every attorney who has used both Claude models eventually forms an intuition about which one to reach for. Most of those intuitions are wrong. The common defaults are "Opus for everything important" and "Sonnet to save money," and both get the decision wrong in at least half the cases.

The actual question is not about cost. The API cost difference between Sonnet 4.6 and Opus 4.8 is so small at typical attorney use volumes that it should not factor into any task-level decision. The question is about output quality on specific task types. The answer differs by task in ways that matter for petition work.

The Price Gap Is Not the Point

Here is the math attorneys worry about. Opus 4.8 runs $5 per million input tokens and $25 per million output tokens. Sonnet 4.6 runs $3 input and $15 output. That is 67% more expensive on output. It sounds significant until you work through actual volumes.

An attorney drafting 20 petition sections per month, each generating roughly 800 tokens of output, is spending $0.24/month on Sonnet or $0.40/month on Opus. The difference is sixteen cents.

The real pricing decision is the Claude subscription tier. Claude Max, which gives access to both Opus and Sonnet, runs $100-$200 per month depending on usage limits. Claude Team, which includes Sonnet but not Opus for sustained petition-drafting volumes, runs $30 per seat. That $70-$170 monthly gap is where the cost question actually lives. And for attorneys who draft two or three EB-1A petitions a month, that gap is typically worth it.

The conclusion: stop making task-level decisions based on per-token costs. Make the subscription decision once, then pick the model based on what the task requires.

What Opus Does Better: Sustained Legal Reasoning

The performance gap between Opus and Sonnet is not about raw intelligence on a benchmark. It is about what happens to argument coherence over 300-400 words of output.

An EB-1A criterion section has a specific structural requirement: it must establish the legal standard (as defined in the USCIS Policy Manual, Vol. 6, Part F), map the beneficiary's evidence to that standard with specificity, and anticipate the most common RFE grounds for that criterion. Done correctly, a Criterion 5 argument runs 400-600 words and keeps each of those elements in coherent relationship throughout. The argument at paragraph five must still serve the framing established in paragraph one.

Sonnet handles this structure reliably up to around 250-300 words. After that, the argument tends to drift. The exhibit citations become less specific. The comparative framing ("placed this work in the top percentile of the field") starts to drop out. The argument closes on a generic assertion rather than on the specific "very top of the field" language that corresponds to the legal standard.

Opus maintains the structural integrity longer. The comparative framing persists across the full argument. The exhibit citations stay specific. The regulatory language appears where the argument requires it, not just in the opening paragraph.

This is not a subtle difference in polished drafts. It shows up as edit time. Opus criterion sections typically need one round of attorney editing for factual accuracy and exhibit numbering. Sonnet sections often need a second pass to restore the legal framing that eroded in the second half of the argument.

The Kazarian (Kazarian v. USCIS, 596 F.3d 1115, 9th Cir. 2010) Step 2 totality argument makes this even more pronounced. Step 2 requires three interlocking components: field definition, comparative standing against a competitive population, and application of the "very top" standard. Sonnet sections frequently collapse components two and three, producing what reads like a restatement of criterion satisfaction rather than a totality-of-the-evidence argument. That structural failure is a known RFE trigger.

Expert letter briefs have the same issue at higher word counts. A full expert letter brief runs 800-1,200 words across multiple sections. Opus maintains coherence across that length. Sonnet does not degrade catastrophically, but it requires more reconstruction of the through-line argument.

For attorneys using Immigration Copilot's EB-1A petition guide, the criterion-by-criterion analysis there maps directly to the task types where the model selection matters most.

What Sonnet Does Just As Well

The majority of an immigration attorney's AI workload does not require sustained legal reasoning. It requires information extraction, organization, and short-form drafting. Sonnet is fully adequate for all of it.

Client email drafts do not require comparative legal framing. Neither do evidence checklists, exhibit indexes, or document summaries. These tasks ask the model to organize information you provide, not to construct a legal argument from that information. Sonnet's output on a "summarize this document" or "create an evidence checklist for these criteria" task is indistinguishable from Opus output to a practicing attorney.

Client intake summaries fall in the same category. You are giving the model a set of facts and asking it to organize them. The output quality ceiling is set by the information you provide, not by the model's reasoning capacity.

Cover letter openings and short (150-200 word) criterion sections also fall here. A brief criterion section that establishes the legal standard and cites two exhibits does not require the multi-step coherence that distinguishes Opus. Sonnet produces a draft that requires comparable editing time.

The practical implication: if you spend 60% of your AI time on intake organization, evidence checklists, and client correspondence, Sonnet handles that 60% at equivalent quality. Reach for Opus when the specific task is on the list below.

The 12-Task Decision Matrix

This table covers the most common AI-assisted tasks in an EB-1A and O-1 practice. The recommendations are based on where output quality differences are measurable, not theoretical.

Task	Recommended Model	Reason
EB-1A criterion section (400+ words)	Opus 4.8	Sustained legal reasoning; comparative framing persists across full length
Kazarian Step 2 totality argument	Opus 4.8	Multi-component structure (field definition, comparative standing, "very top" standard)
Expert letter brief (full document)	Opus 4.8	Multi-section coherence across 1,000+ words
O-1A criterion analysis	Opus 4.8	Complex benefit-of-doubt reasoning, field-specific competitive standing
RFE response strategy	Opus 4.8	Identifies which evidence to emphasize and which RFE grounds to preempt
Short criterion section (150-200 words)	Sonnet 4.6	Adequate quality; reasoning depth not required at this length
Client email draft	Sonnet 4.6	Conversational quality; no sustained legal reasoning
Evidence checklist generation	Sonnet 4.6	Structured list task; information organization, not argument
Document summary	Sonnet 4.6	Extraction task, not a reasoning task
Exhibit index	Sonnet 4.6	Clerical organization; model reasoning capacity irrelevant
Client intake summary	Sonnet 4.6	Information organization from provided facts
Cover letter outline	Either	Run both, keep whichever needs less editing for your style

The cover letter case is the most honest entry in the table. The task sits near the reasoning threshold where both models produce comparable output. If you are on a Max plan with both models available, run the same prompt on each for the first few clients, then pick the default that consistently produces drafts closer to your finished style.

The AI prompts guide for immigration attorneys includes task-specific prompts calibrated for both models. The prompts for Criterion 5, Kazarian Step 2, and expert letter briefs were developed with Opus output quality in mind.

How Immigration Copilot Uses Both Models

The platform this article is published on makes real-world use of both Claude tiers in its petition generation pipeline, and the split is instructive.

Immigration Copilot uses Opus for petition generation and Haiku for document classification. Haiku sits below Sonnet on the model tier: it is the fastest, cheapest, and least capable of the three. That is the right choice for document classification, which is a pattern-matching task. Given a document and a set of immigration criteria, the model assigns labels. That task does not require deep reasoning. It requires speed, reliability, and low cost across hundreds of documents per case.

Petition generation is the opposite. The platform generates 20-60 page petition letters section by section, with each section requiring legally coherent argument grounded in specific uploaded exhibits. That is exactly the use case where Opus earns its cost premium. The quality difference is not marginal at that output volume.

The three-tier architecture (Opus for generation, Haiku for classification, and Sonnet available for intermediate tasks) reflects the same logic this decision matrix recommends for direct attorney use. Match the model tier to the reasoning requirements of the specific task.

For context on the broader EB-1A adjudication environment this petition work addresses, the AAO decisions analysis for 2024-2025 covers recent patterns in how USCIS officers evaluate criterion evidence.

Extended Thinking: When to Use It

Opus 4.8 supports extended thinking mode, a setting that produces a visible reasoning chain before the final response. Sonnet 4.6 does not. For most petition tasks, you do not need it. For Kazarian Step 2, it is worth knowing about.

Extended thinking mode causes Opus to work through its reasoning process before generating the final answer. You see the intermediate steps. The output takes 30-60 seconds longer and costs more tokens. In return, the reasoning chain is more thorough and the final argument is better constructed.

The case for using it is narrow but real. A Kazarian Step 2 totality argument requires the model to hold several moving pieces simultaneously: the definition of the relevant field, the competitive population of practitioners in that field, the beneficiary's standing in that population, the specific evidence that demonstrates "very top" placement, and the framing of all of that in a way that addresses the most common grounds for Step 2 denial. That is a genuinely complex reasoning chain. Extended thinking mode produces better results on it than standard Opus output.

Two practical notes. First, run extended thinking only on the Step 2 argument itself, not on supporting criterion sections. The latency and cost increase is not justified for tasks that do not require the full reasoning chain. Second, the thinking output (the reasoning steps you can see) is useful for attorney review. It shows how the model weighted the evidence and which comparisons it considered. That transparency helps attorneys catch reasoning errors before they appear in the filed document.

For other petition tasks on the Opus list above (criterion sections, expert letter briefs, RFE strategy), standard Opus mode produces adequate results without the added latency.

Setting Both Models Up in Claude Projects

If you are on a Claude Max plan with access to both models, the most practical setup is two separate Projects: one configured for Opus tasks, one for Sonnet tasks.

The Opus project should contain your full EB-1A or O-1 system prompt: the legal standard from 8 CFR 204.5(h)(3), the Kazarian two-step framework, your output format requirements (exhibit citation format, section length targets), and the constraint that the model should not state facts not provided in the conversation. This system prompt applies to every criterion section, Step 2 argument, and expert letter brief you draft in that project.

The Sonnet project can use a shorter system prompt for the task types it handles: client-facing tone guidelines, checklist format preferences, intake summary structure. These tasks do not require the full legal context that Opus needs.

Switch between projects based on the task, not the client. A single EB-1A case will generate tasks for both projects: Opus for the criterion sections and Step 2, Sonnet for the evidence checklist and client status email.

For the full setup workflow including system prompt text, see the Claude for immigration lawyers guide. The prompt library in the AI prompts guide includes prompts that work in both project configurations.

If you want a platform that handles the model routing automatically (using the right tier for each petition task without requiring manual project switching), Immigration Copilot's petition generation pipeline does that by default. The generation workflow assigns Opus to criterion sections and Step 2 arguments, Haiku to document classification, and routes each task to the appropriate model without attorney configuration.

For the full picture of AI tools available for immigration practice, the AI in legal practice hub covers ethics, BAA requirements, and workflow fit across tools.

See how Immigration Copilot handles petition generation