What AI Models Does Harvey Use?

Our benchmark that measures the ability of LLMs to complete real-world legal tasks. It evaluates both general-purpose LLMs and Harvey’s specialized agentic systems using detailed, lawyer-designed rubrics that score answer quality (accuracy, completeness, legal reasoning) and source reliability (verifiable citations to legal documents).

• This approach ensures the models are judged not just on linguistic output, but on their ability to perform trusted, billable legal work with traceable, high-fidelity results. To learn more, read our blog article on Introducing BigLaw Bench.

In a test environment, we integrate the model into key product surfaces–Assistant, Vault, and Workflows–and measure system performance through both human preference and traditional machine learning metrics (accuracy, precision, recall, hallucination rate, latency, and more) over product datasets.

Researchers test the model in open-ended, less predictable ways to uncover negative features such as toxicity or bias, as well as identify major, sudden improvements in reasoning ability.

Generative AI refers to a class of artificial intelligence systems designed to generate new content, ideas, or solutions based on patterns and data it has learned from. It uses machine learning models to create text, images, music, and even code or designs, mimicking human creativity. Unlike traditional AI, which focuses on recognizing patterns or making decisions based on existing data, generative AI can produce novel outputs, often by learning from vast datasets.

When you input information into Harvey, or when Harvey generates a response, it's processed in tokens. A token is the smallest unit that a model can understand. Tokens have specific statistical properties that make them better to use than just using words.

• Try using OpenAI’s tokenizer for an example of how words map to tokens.
• For more technical information, read Understanding Context Windows and Token Limits

• Think of a "context window" as the amount of information an AI model can 'see' and process at one time. When you're working with Harvey on a document, the context window determines how much of that document the model can consider to generate its response.
• A larger context window means the model can review more text at once, leading to more comprehensive and accurate analysis, especially for lengthy and complex documents like credit agreements.
• For more technical information, read Understanding Context Windows and Token Limits

Yes, LLMs are trained on data up to a certain point in time, known as a "knowledge cutoff date." This means the model's knowledge of events or information beyond that date may be limited.

• If you select a model via the Model Selector, the knowledge cutoff would depend on the model used in response to a query. Most of the models we support have similar knowledge cutoffs.
• If you’re using "Auto" mode or don’t have a model selector available, the knowledge cutoff is June 2024.
• When you use Harvey, we combine AI's broad knowledge with specific, real-time information from your documents or trusted data sources, to optimize performance.

If you have model selector enabled, you’ll notice we note whether each model is best for complex or general-purpose tasks.

• General-purpose tasks: Summarizing and analyzing documents, writing, ideation, and more everyday work.
• Complex tasks: Multi-step reasoning, long-form drafting, complex workflows, and more.

  • Models tagged for “complex tasks” are reasoning models that take more time to think before answering — this usually lets them perform better on tasks that are at the edge of model capabilities. They may be slower to process, and sometimes overthink the easy things.

If you’re not sure where your task falls, choose Auto — Harvey will line up your task with the best model for the job.

Harvey is built by and for the world's best legal and financial companies, and data security is paramount. We engineer our platform with robust privacy and security measures. Sensitive data is handled with the utmost confidentiality and is not used to train the underlying public AI models, nor is it shared externally without your explicit permission. We prioritize practical solutions that honor established practices while using technology to enhance professional work.

For more information, visit our Security Portal.

• While AI models are incredibly powerful, they’re designed to be sophisticated reasoning tools. Sometimes, models might provide a very comprehensive response that seems to "overthink" a simple query. This occurs because these models are built for deep reasoning on hard problems.
• If you encounter an unexpected response, consider rephrasing your question or, if available for the task, trying a different model that is optimized for more succinct answers.

Updates to the models powering Harvey occur regularly, aimed at continuous improvement and enhancing your experience without requiring you to manage complex technical configurations.

Although most users will benefit from Harvey's default selection, power users might like the option to choose themselves due to personal model preferences. You may find that you prefer the results or tone of certain models depending on the use case.

Not at this time, our multi-model use makes it challenging to break this down clearly per query, however this request is on our development list for future consideration.

Harvey specializes in complex professional tasks, particularly in legal and tax workflows.

No, Harvey is built so you can work with simple, clear instructions and minimizes the need for detailed prompt engineering. We also provide example libraries and features to automatically refine prompts for better outputs.

Think of Harvey as a digital associate—a fast and effective thought partner and generator of first drafts. While the output requires verification, similar to the work of a junior colleague, it saves significant time and enhances overall quality. Harvey also allows you to save favorite prompts for future use, expediting workflows and saving you time.

For more on getting the most out of your prompts, visit our Prompt Writing article.

Usage is measured relatively and is rarely, if ever, exceeded. If Harvey observes that a Customer’s monthly usage is materially disproportionate compared to other similarly situated customers, Harvey will notify the Customer. If, after 48 hours of such notice, the Customer’s usage remains outside typical usage patterns, Harvey may reasonably limit the Customer’s access to the Service to help maintain overall service quality.

Hallucinations are a type of error that occurs when LLMs (Large Language Models) generate inaccurate or fabricated information. This can happen because LLMs are trained on large and diverse corpora of text, but may lack sufficient domain knowledge or logical reasoning.

Harvey minimizes hallucinations by using domain-specific models and knowledge bases:

• Domain-Specific Models: These models are trained on large datasets of specialized legal documents, capturing the nuances and complexities of the law. This narrows the gap between general-purpose models, which lack domain expertise, and human experts with specialized knowledge.
• Knowledge Bases: Many of Harvey’s tools use domain-specific resources, such as statutes, case law, and legal ontologies, to ground responses in authoritative sources. For legal tasks, Harvey references selected case law and statutes, and when documents are uploaded, the output will include citations from the user-provided materials.

We invest in several areas that go beyond training a single model:

• Composable systems: Harvey combines many small model “calls” together. This means a single query might use dozens—or even hundreds—of steps behind the scenes to give you a complete answer with supporting references.
• Multi-model strategy: We partner with more than one AI provider. Each model has different strengths, so leveraging them all makes our results more accurate and reduces the need for heavy custom training.
• Orchestration: As our systems get more complex, we route your request to the right workflow. This helps ensure you get the best possible result without extra effort.
• Workflows: We’re creating step-by-step processes that guide you through legal work from start to finish.

Harvey uses domain expertise to convert professional processes into high-quality AI agents that produce expert-quality work product:

• Each request is routed to a cascading series of LLMs tuned for legal synthesis, RAG systems that incorporate public or user-provided data, and powerful reasoning models like o1 that help orchestrate the work end to end.
• This allows Harvey to both reduce hallucinations and improve response output quality.
• You can read about Open AI’s explanation of RAG vs. fine-tuning here.

Our system is designed to break up a larger task into smaller tasks that AI models are more likely to execute correctly. We created a “Partner”-like model to create a plan and delegate each subtask to the right model. The type of work being completed determines the model that is used.

The Details

We build our models to be similar to how a partners coordinate complex, multi-step tasks—delegating the work in a specific order to specific associates.

How It Works

1. Our Partner model creates a plan to solve the overall tasks and identify which subsystems to delegate portions of each task to.
2. Based on these assignments, our subsystems will receive the necessary context from the Partner model.
3. The subsystem the performs the task to the best of its abilities.
4. The output is sent back to the Partner model.
5. The Partner model synthesizes the output and produces the final result.

Example (Case Law)

Let’s say you ran this query: “Can the pass through defense be used to defeat class cert in nd cal.” The Partner model might generate the following list of tasks:

1. A list of sub questions that need to be answered to answer the original question.
2. A list of case law searches to answer each sub question and examples of what the subsystems should look for.
3. A simple plan to answer the subquestions (“summarize these cases and highlight these things for each subquestion”).
4. Synthesize the results into a final answer.

Example (Drafting)

Let’s say you ran this query: “Draft a client alert on the attached EU AI Act of 2024” with an uploaded copy of the EU AI Act of 2024. The Partner model might generate the following list of tasks:

1. A list of sub questions that need to be answered to get all the relevant information from the provided document.
2. A simple plan to answer the subquestions (“summarize these details and extract key information from these things for each subquestion”).
3. An outline of the client alert, to be filled in with information collected above.
4. Synthesize the results into a final answer, following the outline.

Currently GPT-5 thinking is the default and only version for selection in Harvey's Model Selector. We’re in the process of quality testing GPT-5, which will be incorporated and optimized in Auto mode once testing is complete. For the best performance, we recommend using Auto mode, which is designed to deliver the highest-level reasoning and adapt as we continue improvements.