Now Available in Lito: Claude Opus 5
Summary
Claude Opus 5 is now available in Lito. Litera's Legal Knowledge Engineering team evaluated the model's behavior on legal work in Lito, including complex document analysis, source-grounded reasoning, drafting, summarization, extraction, and classification tasks. The evaluation found that Opus 5 performed reliably on judgment-heavy work, with generally specific source citations and coherent reasoning where conclusions depended on subtle document language rather than explicit statements.
The model is not self-validating. Its principal limitations are calibration-related: it can extend beyond the requested scope, provide more detail than the task requires, and occasionally apply a broader interpretation than intended in extraction or classification work. Lawyers get more reliable results when they specify the legal question, scope, jurisdiction, source materials, and desired output format, then review the output against the underlying documents.
TL;DR
Claude Opus 5 is now available in Lito and performed well in Litera's qualitative evaluation of complex legal tasks. It is a strong option across the full range of legal work, from quick lookups to the hardest judgment calls, when lawyers provide precise instructions, ground the work in source documents, and review the output in proportion to the task's stakes.
Assessing Legal AI for Work That Requires Judgment
Claude Opus 5 is now available in Lito. For lawyers, the relevant question is not where a model sits in a general-purpose benchmark ranking. It is whether the model can be used responsibly on work that carries legal and commercial consequences.
Litera's Legal Knowledge Engineering team evaluated Claude Opus 5 in Lito across legal tasks that require more than locating an obvious answer in a document. The work included analysis of subtle document language, source-grounded reasoning, drafting, summarization, extraction, and classification.
The results were encouraging, with an important qualification. Opus 5 performed reliably on complex legal work. Still, like any large language model, it benefits from clearly defined instructions, relevant source material, and review proportionate to the stakes of the task.
Is Claude Opus 5 Accurate Enough for Legal Work?
A reliable evaluation is not a reason to skip review. For legal work, the relevant standard is not whether an answer sounds plausible. It is whether the answer is grounded in the relevant materials, responds to the question asked, reflects the appropriate jurisdiction and legal context, and has been reviewed with appropriate care.
For work with legal or commercial significance, the lawyer remains responsible for defining the task, evaluating the output, and confirming conclusions against the underlying documents.
How Opus 5 Performed in Legal Work
Reliable on complex, judgment-heavy tasks
Opus 5 performed well not only on straightforward document questions, but also on tasks requiring a conclusion from nuanced language. In the Legal Knowledge Engineering team's evaluation, the model generally maintained coherent reasoning and returned source citations that were specific enough to support review of the underlying material.
Where lawyers need to set clear boundaries
Opus 5 tends to be generous with scope. It may add adjacent context, advisory observations, or drafting notes beyond what the user requested. In some situations, that additional material may be useful. In others, it can make the answer less efficient to review or produce an output that is not suitable for immediate use.
The same pattern can appear in formatting. A task that calls for a direct answer may produce a comparison table, additional explanation, or an extra caveat unless the user specifies the preferred format and response limits.
How to get more concise output
The Legal Knowledge Engineering team also observed that Opus 5 often produced responses that were longer than the task required, even when the substance was accurate. This is separate from the question of scope: a response may answer the right question but take two or three times as much space as is useful for the task.
Directional prompts with clear instruction can materially improve the result. For example:
Provide a concise answer of no more than five bullets. Do not include background analysis, drafting notes, caveats, or introductory commentary unless they are necessary to answer the question.
The point is not to make the model less thorough where thoroughness is needed. It is to align the output with the lawyer's intended use, whether that is a quick issue list, a client-ready summary, or a deeper analysis.
How to Prompt Legal AI for More Reliable Results
The same habits improve results across models and legal tasks.
Define the legal question precisely
Ambiguous instructions can lead to an answer that is technically responsive but does not reflect the user's intended interpretation. Use the relevant legal terminology, define any terms that may have multiple meanings, and state the decision or conclusion the output should support.
Set the scope, including exclusions
Explain what to include and identify what to leave out. A narrowly defined request is especially useful for extraction, classification, and drafting tasks, where the model may otherwise include adjacent issues or provide additional analysis that is not needed.
Identify the jurisdiction
If jurisdiction is not stated, a model may default to U.S. law without expressly identifying that assumption. State the applicable jurisdiction, governing law, and any other relevant legal framework when the analysis depends on them.
Ground the task in current sources
Where a question depends on the current state of the law, provide the relevant authority or source material rather than relying on the model's recalled knowledge. Similarly, when reviewing agreements or transaction materials, use the source documents as the basis for the request.
Specify the preferred format and length
A model cannot infer whether the user needs a one-paragraph executive summary, a draft clause, or an issue list with citations. State the intended format, desired level of detail, and any length limit.
How Lawyers Should Review AI-Generated Output
AI-assisted work still requires review. The appropriate level of review should reflect the output's complexity, materiality, and downstream use.
For document review tasks, lawyers should:
- Confirm that the output addresses the precise question asked.
- Check source citations and quoted language against the underlying documents.
- Validate the completeness of extraction lists, especially when provisions are repetitive or nearly identical.
- Confirm that each extracted or classified item meets the category defined in the prompt.
- Review nuanced legal conclusions to ensure they are grounded in the relevant language and authority.
- Confirm jurisdictional assumptions and use current, supplied sources where the task depends on current law.
The first response is often best treated as a well-developed draft. Follow-up prompts can narrow the scope, correct an assumption, request a different format, or ask the model to identify and address gaps. That iterative approach is often more reliable than expecting a single prompt to resolve every ambiguity in an open-ended legal task.
Claude Opus 5 Is Now Available in Lito
For lawyers using Lito, Claude Opus 5 is a strong option for a broad range of legal tasks, including complex document analysis, drafting, and summarization. Its main limitation is not an inability to handle difficult work, but the need for calibration: without clear instructions, it may provide more analysis, context, or length than the task requires.
The most effective approach is straightforward. Define the task precisely, provide the documents and authorities that should ground the answer, specify the scope and format, and review the result with the same professional judgment the work requires without AI.
Learn more about Lito or request a demo to see how Litera AI can support legal work across the transaction lifecycle.
These observations reflect a focused qualitative evaluation of Claude Opus 5's default behavior in Lito by Litera's Legal Knowledge Engineering team at a point in time. Model behavior evolves, and individual results will vary by prompts and documents.
