What AI model does it use?

What AI model does Harvey use?

Harvey, the AI tool for law firms, is still often described as "built on OpenAI". That was true at the start. Its own blog and release notes, read on September 25, 2026, describe something broader: models from three big labs, a selector lawyers can use, and a model Harvey trained itself.

From OpenAI to three labs

The change came on May 13, 2025, when Harvey announced "the integration of Anthropic and Google models to supplement our existing OpenAI models in the Harvey platform." In March 2026 Niko Grupen explained the reasoning in a post titled "Why Harvey is multi-model by design", with examples: "Claude Opus 4.6, for example, excels at agentic reasoning and deep multi-step legal analysis, while GPT-5.2 pairs strong analytical abilities with capability awareness."

The same post says speed matters on some features. For high-volume ones such as Vault (where firms store and bulk-analyze documents), Harvey found that "models like Sonnet 4.6 and Gemini 3 Flash deliver strong performance with a significantly lower latency footprint."

The model selector

Lawyers aren't stuck with Harvey's choice. "If an individual practitioner works most effectively with a particular model, they can select it directly in our model selector," Grupen wrote. Workspace administrators decide which models their firm can see, and Harvey can also "route work to an alternative model" if a provider has an outage.

The newest arrivals are in the release notes. On September 18, 2026, "Fable 5.1 is now available as a model provider in Harvey, giving workspaces that opt in access through the model selector", and GPT-6 Astra arrived the same day for workspaces that request it. So both labs' top models are there, each switched on by the firm rather than for everyone.

A model Harvey trained itself

In August 2026, Harvey and a company called Applied Compute described training a model for Review Table, the Harvey feature that pulls answers out of many documents into a grid. They started from GLM 5.2, an open model from Z.ai, and on Harvey's own benchmark the result "reached 0.903 on our benchmark, compared with 0.867 for Fable 5 and 0.857 for GPT-5.6-Sol."

That's Harvey's own test on one task, and the post doesn't say whether customers are using the model yet. Still, a legal AI company beating the big labs' flagships on its own feature with an open base model is a different picture from "Harvey runs on GPT".

Who else handles the data

Harvey's subprocessor list, dated September 21, 2026, fills in the rest. Alongside OpenAI, Microsoft, Google Cloud, AWS and Anthropic as AI providers, it lists Mistral, ElevenLabs for voice, DeepL for translation, and Baseten and Fireworks.ai, two companies that host open models.

What Harvey doesn't say

There's no published default. Which model answers when a lawyer doesn't choose one is Harvey's call, and it isn't listed anywhere I could find.