GPT-5.6 Sol vs. Terra: Which Model Should You Use for Legal Work in Lito?
Summary
GPT-5.6 Sol and GPT-5.6 Terra are now available in Lito, the award-winning Legal AI agent, and they're built for different kinds of legal work. Sol is OpenAI's flagship model, optimized for complex, high-stakes tasks. Terra is designed for focused, well-scoped everyday work. Litera's Legal Knowledge Engineering team, composed of former practicing lawyers across multiple jurisdictions and practice areas, evaluated both on real legal tasks in Lito. Here's what they found and how to choose.
TL;DR
- Sol runs deeper on complex tasks and is the stronger choice when thoroughness and full analysis matter
- Terra is faster and more concise, built for focused, well-defined work where speed matters
- Both are in Lito now, and the model you reach for should follow the task
In This Article
- What Are GPT-5.6 Sol and Terra?
- Sol: Strengths and Trade-Offs
- Terra: Strengths and Trade-Offs
- Which GPT Model Should You Use?
- How to Get Better Results from Sol or Terra
- What These Models Won't Do for You
- Frequently Asked Questions
Many capable models sitting in the same platform sounds like a good problem to have. Most lawyers who open Lito will pick one and stick with it. But Sol and Terra aren't interchangeable, and on complex legal work, the difference is noticeable.
Litera's LKE team evaluated each model on GPT-5.6 legal work across the tasks that define transactional and deal practice. Here's what they found.
What Are GPT-5.6 Sol and Terra?
GPT-5.6 is OpenAI's newest model generation. Two members of that family are now in Lito: Sol, which OpenAI positions as its flagship model built for the hardest, highest-stakes tasks, and Terra, which OpenAI describes as a balanced model for everyday work.
Some tasks need a model that digs past the question and brings back the reasoning behind its conclusions. Others need a model that takes a precise ask and returns a tight answer quickly. Sol is for complex, open-ended work, and Terra is for focused, well-scoped tasks.
Sol: Strengths and Trade-Offs
Sol is the more thorough of the two. On tasks with several moving parts, it's more likely to catch all of them, surfacing adjacent provisions the prompt didn't explicitly request and carrying the reasoning further than a narrower model would.
That instinct shows up most clearly on the hardest work. On contested drafting points, Sol gives you the answer with the reasoning behind it and the qualifications that matter. On open-ended research tasks and memos drafted from scratch, evaluators found Sol's output closer to what a partner would want to see.
But that same thoroughness means Sol's default output leans generous. On a straightforward, well-defined task, it will often add material that wasn't called for: context, adjacent sections, additional framing. OpenAI describes the GPT-5.6 family as more concise than the prior generation, and Sol is economical relative to its capability level. On focused tasks, extra material lengthens the answer without improving it, which means more time spent filtering.
When the task is complex and the output needs to be thorough, Sol is the stronger choice.
Terra: Strengths and Trade-Offs
Terra is faster and stays closer to the ask. On focused, well-defined tasks, it returns output more promptly and doesn't pad the answer.
In the LKE evaluations, Terra rated slightly lower than Sol on accuracy overall, and the gap was narrow. On the tasks Terra is designed for — targeted clause extraction, quick lookups, high-volume routine work — the difference in output quality is not material. You get what you asked for, returned quickly.
Now and then, Terra drops something the prompt didn't explicitly demand, but that would have been useful: a section reference, a related provision that adds context. A follow-up prompt usually fixes it. On complex work, the extra exchanges start to cost more time than the speed advantage is worth.
When the task is focused and the scope is clear, Terra is the stronger choice.
Which GPT Model Should You Use?
The question isn't really "which model is better." It's "which model fits this task." Here's how that plays out across the work LKE evaluated.
| Reach for Sol when… | Reach for Terra when… |
|---|---|
| The task involves contested drafting points or partner-level judgment calls | The task is focused and the scope is clearly defined |
| You need adjacent provisions surfaced without having to ask | The task calls for speed and a focused answer |
| You're drafting from scratch: memos, complex research, multi-part analysis | You're doing targeted extraction or quick lookups |
| A missed qualification would create risk | You expect to refine the output with follow-up prompts |
| The matter is jurisdictionally complex | The task sits in familiar territory |
Sol and Terra serve different purposes, and the choice between them depends on the task. Using Sol on a simple extraction task makes the answer longer, not better. The model you reach for should fit the work in front of you.
How to Get Better Results from Sol or Terra
How you use either model matters as much as which one you pick. The LKE team identified four habits that apply to both Sol and Terra and consistently produce better output.
- Be precise about what you want. Some answers that look like misses are just a mismatch between what the user expected and how the model read the prompt. Legal terminology can be ambiguous and a term with multiple accepted meanings will sometimes produce a response anchored on the wrong one. The same goes for output format: absent a stated preference, either model will make its own call on structure. The more precisely you state your expectations, the more reliably the model will meet them.
- Treat the first response as a starting point. On complex or open-ended tasks like research, multi-part analysis, or a memo drafted from scratch, plan for back-and-forth rather than a finished answer in one pass. One pass is rarely enough on work like this. Going back in with a tighter prompt is part of the process.
- State your jurisdiction. When a question doesn't specify one, both models can default toward U.S. law in subtle ways. How a date is read for an agreement governed by non-U.S. law is one example. LKE found that Sol and Terra flagged jurisdictional issues at times, and the problem still appeared in testing. Stating your jurisdiction upfront prevents a category of errors.
- Verify exhaustive lists against the source. On large-scale extraction tasks like risk factor analysis in a Form 10-K, a near-complete list can read as complete. A single near-duplicate item can drop out silently. On tasks like these, be sure to cross-reference the output against the source document.
What These Models Won't Do for You
Sol and Terra both perform well on legal work. They're also large language models (LLMs), and the same limitations that apply to any LLM apply here.
Assertions on nuanced points of law should be verified against the cited source. References too general to be traceable should be treated with caution. On matters where a wrong answer creates professional or client risk, the output is a starting point for legal judgment.
Litera is trusted Legal AI that best unifies the practice and business of law. Lito is the Legal AI agent that delivers intelligence for both sides of the firm. When accuracy matters most, proven deterministic engines power precise outputs. When the work calls for flexibility, lawyer-trained skills and custom workflows built on the latest foundational models — like Sol and Terra — are available. Every matter and client interaction feeds one data layer, so your own work becomes intelligence that surfaces the next opportunity and wins the next matter.
Knowing the limits of these models is what will allow you to use them well.
Frequently Asked Questions
Are GPT-5.6 Sol and Terra available in Lito now?
Yes. GPT-5.6 Sol and GPT-5.6 Terra are both available in Lito now. You can switch between them based on the task.
How accurate are Sol and Terra on legal tasks?
In LKE evaluations, Sol ranked slightly behind Fable (OpenAI's most advanced GPT-5.6 model) on accuracy and ahead of Terra. The gap between them was narrow and consistent across the tasks Terra is designed for. Both models perform well on legal work, and the choice between them is about depth and scope rather than accuracy.
Which model is better for contract review?
It depends on the nature of the review. For targeted extraction: pulling specific clauses, checking defined terms, and identifying particular provisions, Terra's conciseness and speed are an advantage. For contract review involving contested language, jurisdictional complexity, or multi-party risk assessment, Sol's analytical depth is the stronger choice.
Can I use the same prompts for both Sol and Terra?
You can, and precision matters more with Terra. Because Terra stays closer to the explicit ask, a vague prompt leaves more on the table. Sol returns broader output on the same prompt, though on simple tasks that means more to filter. Tighter prompts produce better results with either model.
How do OpenAI's GPT-5.6 models compare for legal drafting?
Both are more concise than prior-generation reasoning models. For drafting from scratch: client memos, complex research, multi-part documents, Sol's thoroughness and tendency to carry reasoning further is the stronger choice. For shorter, well-defined drafting tasks, Terra gets you to a working draft faster with less filtering required.
Both GPT-5.6 Sol and Terra are solid choices for legal work in Lito. Reach for Sol when the work is complex and completeness matters. Reach for Terra when the task is focused and speed is useful. Both have been evaluated on legal tasks by lawyers.
See how Lito works
Meet us at ILTA
