Claude Opus 4.7 vs Sonnet 4.6 When to Use Which

This comparison concerns two earlier Claude models, not the current model lineup. The share of a firm’s work suitable for Sonnet 4.6 is unknown without a controlled evaluation of its actual tasks. Compare material errors, source fidelity, reviewer time and completed-task cost before assigning work to either model. Anthropic’s pricing documentation lists standard first-party API rates of $5/$25 per million input/output tokens for Opus 4.7 and $3/$15 for Sonnet 4.6; these are token rates, not completed legal-task costs. For the newer release, see Sonnet 5.5 for Lawyers.

What separates Opus 4.7 from Sonnet 4.6 in actual legal work

Confirm model availability, context limits, effort controls and deployment features in the current documentation for the exact API, account or external platform. Features of an application or agent environment are not automatically model guarantees.

A feature list does not establish legal accuracy. Verify the selected model’s documented controls, then test them on the intended task; software and science benchmark scores do not measure legal reliability.

A feature list does not establish legal accuracy. Verify the selected model’s documented controls, then test them on the intended task; software and science benchmark scores do not measure legal reliability.

The practical differentiation in legal work shows up in three places:

Calibration on legal questions needs a controlled evaluation. Include jurisdiction-specific questions, missing authorities and plausible but incorrect premises. Record material errors and whether the model expresses uncertainty appropriately; no comparative legal-calibration result is established here.

Matter continuity depends on the deployed product, tools and storage architecture. Test what persists across sessions, who can access it, and how it can be removed. This article does not establish that one model alone supplies safe matter memory or a fixed amount of recovered time.

For complex tasks, compare outputs against a defined source packet and reviewer criteria. More effort, a longer answer or a higher token rate does not establish a correct legal conclusion.

Routing is an operational decision. Let measured review effort, material errors and completed-task cost drive the choice; no evidence here establishes how most firms route their queries.

The routing matrix: which legal tasks get Opus vs Sonnet

Candidate evaluation for Opus 4.7:

Test a bounded diligence task, a difficult source-grounded argument and a permitted scanned-document example. Compare them with the same tasks on the baseline model. A model choice does not establish privilege protection or remove independent review.

Candidate evaluation for Sonnet 4.6:

Test a synthetic contract against an explicit checklist, a bounded first draft, and document classification against known labels. Common templates do not guarantee sufficient accuracy; retain review and rejection criteria.

Mixed routing (use both within a single workflow):

A proposed mixed workflow can generate a draft with one model and review selected issues with another. Test whether the second pass finds material errors rather than repeating them. Neither model’s agreement is independent verification.

Set the routing split from the firm’s evaluation results; this article has no evidence establishing a typical percentage of legal work for either model.

Pricing reality: what the routing decision saves

Hypothetical cost illustration: assume 100 million tokens per month, 70% input and 30% output, at the standard token rates above. A 70% Sonnet / 30% Opus routing split is an invented arithmetic assumption, not an observed legal-workload share. The example excludes seats, taxes, caching, tool charges, tokenization differences, setup and human review.

Hypothetical 100% Opus token bill: 70M input × $5/M + 30M output × $25/M = $1,100/month, or $13,200/year. Application seats and API usage are separate purchases; no seat price is added here.

Hypothetical 70% Sonnet / 30% Opus token bill: Sonnet $462/month plus Opus $330/month = $792/month, or $9,504/year. This assumes the same input/output mix for each model and an unchanged number of billable tokens.

The difference between those hypothetical token bills is $3,696/year. It is not a measured saving or forecast for a firm. A change in token use, review effort, task success or deployment pricing can alter or reverse the result.

Measure total workflow cost, including preparation, retries, corrections and reviewer time. A lower token bill is useful only if the completed work meets the same acceptance criteria.

Tokenization and platform pricing can change the comparison. Price the same source packet through the actual deployment rather than assuming equal token counts or projecting a firm-wide saving.

Operational patterns: how firms actually deploy both models

Three proposed deployment patterns to evaluate:

Pattern 1: Default-Opus with Sonnet fallback for bulk work.

In a proposed default-Opus workflow, the team assigns a defined bulk task to Sonnet only after that task meets the same acceptance criteria. Confirm model availability and access on the chosen deployment.

Treat this as an implementation option. Test whether the team can operate and review it consistently before adopting it.

Pattern 2: Default-Sonnet with Opus escalation.

In a proposed default-Sonnet workflow, define explicit escalation and rejection conditions. A higher-cost model still requires verification; high stakes alone do not establish model suitability.

Treat this as an implementation option. Test whether the team can operate and review it consistently before adopting it.

Pattern 3: Workflow routing.

In a proposed workflow-routed deployment, record each task’s model, source packet, result, review and escalation. Evaluate whether the added routing complexity improves the completed work.

Treat this as an implementation option. Test whether the team can operate and review it consistently before adopting it.

Routing can be implemented in an approved application or middleware layer. This article has no survey establishing which architecture most large firms use.

Recommendation by firm profile

Solo and small firms: compare a simple approved workflow with the added routing complexity. Confirm current access and pricing; no universal default model or subscription sufficiency is established here.

Mid-market firms: test one recurring workflow and include the cost of operating the routing layer. No firm-size savings range has been measured here.

Large firms: review ownership, deployment controls and evaluation coverage before introducing routing. No implementation timetable or savings range is established here.

Practice area helps select test cases, not a predetermined model ratio. Compare contract review, diligence, litigation and regulatory tasks separately. Set the routing split from the firm’s evaluation results; this article has no evidence establishing a typical percentage of legal work for either model.

Pick the deployment pattern that matches the firm's operational capability, not the pattern that promises the largest theoretical savings. A pattern the firm can't sustain delivers worse outcomes than a simpler pattern with smaller savings.

Choose routing from the firm’s actual evaluation results. Set the routing split from the firm’s evaluation results; this article has no evidence establishing a typical percentage of legal work for either model. Include material errors and reviewer effort in the decision; this article does not establish first-month payback or universal savings.