Three companies all launched a dedicated legal AI product within four months of each other. What each one actually does differently, explained simply.
To the best of my ability, I have been following the LegalTech AI updates, and after some time I concluded that I have no understanding of the main differences in approaches between the big three: Anthropic, Google and OpenAI. So I decided it was time to write an article for other students like me, trying to picture what the career will look like in the future.
What is the difference between the big three new law AI models?
Before we begin, I am going to be using an analogy throughout to put it simply. Picture a shopping mall, and inside it, two very good specialist shops: Harvey and Legora.
When it was made. Claude for Legal launched on 12 May 2026, months before either of the other two.
How it works. Rather than building one single legal model, Anthropic's approach is a set of role-specific plugins layered onto Claude itself, including an employment counsel plugin, a litigation associate plugin, and even one built for law students. Alongside the plugins, Claude connects directly into the tools lawyers already use daily, including Thomson Reuters' Westlaw and CoCounsel, and is embedded directly into Word, Outlook, Excel and PowerPoint.1
Real firms are already using it on live matters, including Freshfields, Quinn Emanuel and Holland and Knight.2 Anthropic's own associate general counsel has said that outside of software developers, lawyers are the single most active professional group using Claude.1 A big strength of Claude's is its context window, up to one million tokens, allowing it to hold 1,500 to 2,000 pages in a single session, useful for working through long contracts or an entire deal binder at once.3
The analogy. Claude already runs its own large, well known general store in the mall, that's the Claude chatbot people already use directly. On top of that, Claude also connects directly into Harvey and Legora's own shops.
When it was made. Gemini Enterprise for Legal launched a bit later, on 25 August 2026.
How it works. This isn't built around a single powerful model doing the legal reasoning itself, like Claude. It is a platform that connects to the tools firms already use, including Harvey and Legora themselves, through something called MCP, a way for one piece of software to talk directly to another. In practice, this means a lawyer can ask a question inside Gemini, and rather than Gemini answering itself, it will be routed straight to the real work happening inside Harvey or Legora, with the answer linking back to the original source.4 Launch firms named alongside the product include Cleary Gottlieb, Freshfields, Weil, and Williams and Connolly.5
The analogy. This is different to Claude. Claude connects into Harvey and Legora's shops, but once it's there, it still does its own thinking, it's Claude's own knowledge doing the work. Gemini works differently. It understands your question, sends the specialist part of the work to Harvey or Legora, and brings the answer back to you in one place, but the legal reasoning itself happens inside their shop, not Gemini's. Gemini is the one managing the conversation. It isn't the one doing the legal thinking.
When it was made. Astra for Law is the most recent of the three, launching on 17 September 2026.
How it works. It is built around one model, GPT-6 Astra, OpenAI's most powerful, wrapped in a dedicated legal search index covering over 230 million US case law, statute and regulation documents, updated daily and built with the nonprofit behind CourtListener. Access is currently limited to a small number of invited firms through something OpenAI calls Trusted Access, and no pricing has been published yet. Harvey and Legora have both been named as early customers, meaning they will be able to build their own products using Astra for Law too.6
A live update, just five days later. On 22 September 2026, OpenAI announced two further models, GPT-6 Sol and GPT-6 Luna, built to carry much of Astra's strength into faster, cheaper models for higher volume work. Sol is aimed at complex tasks, Luna at simpler, high volume ones like summarising documents.7 Harvey has a consistent track record of adding each new OpenAI model quickly, it did the same with GPT-5, GPT-5.5 and GPT-5.6 Sol before this.8
The analogy. This is different again to both Claude and Gemini. Claude connects into Harvey and Legora's existing shops and does its own thinking while it's there. Gemini connects into their shops and lets them do the thinking. OpenAI didn't connect into an existing shop at all, it built a brand new one, stocked with its own private library. For now, only a small number of invited customers can shop there, and Harvey and Legora are among them.
| Claude for Legal | Gemini Enterprise for Legal | Astra for Law | |
|---|---|---|---|
| Launched | 12 May 2026 | 25 August 2026 | 17 September 2026 |
| Core approach | Plugins layered on Claude, plus direct connections into existing legal tools | A connector platform that routes questions to specialist tools | One model plus a dedicated, proprietary legal search index |
| Does it think, or does it route? | Thinks for itself, even when connected to other tools | Routes the question, the thinking happens in the connected tool | Thinks for itself, using its own index |
| Where it lives | Embedded in Word, Outlook, Excel and PowerPoint | Connects into legal platforms via MCP | Delivered through ChatGPT and Codex |
| Named firms | Freshfields, Quinn Emanuel, Holland and Knight | Cleary Gottlieb, Freshfields, Weil, Williams and Connolly | Selected Am Law 200 firms (Trusted Access) |
| Access right now | Paying customers only, via the Cowork desktop app, plugins added manually. Still a research preview | In preview | Invite only, no public pricing |
| Connected to Harvey and Legora | Yes, directly | Yes, as named launch partners | Yes, as named API customers |
| Most recent update | GPT-6 Sol and Luna announced 22 September 2026, extending Astra to faster, cheaper models |
So, what does each of them actually mean by "legal AI"? Claude decided the general assistant should keep doing the legal thinking itself, reaching outward to specialist tools only when it needs to, though its legal tools still sit behind a real setup step, a separate desktop app, manual installation, and an explicit research preview label. Gemini decided the thinking should stay with the specialist tool, building itself into the doorway instead, with the strongest built in answer for how permissions and confidentiality carry across systems, but it is still in preview. OpenAI decided neither of the above was enough, building its own separate foundation with a frontier model and a genuinely enormous proprietary case law index, though almost nobody outside a small group of invited firms has used it yet.
What ties the three together is more revealing than what separates them. None of them think the answer is doing everything alone, and none of them think the answer is stepping back entirely either. Somewhere between full independence and simply pointing elsewhere is where all three have chosen to sit, they have just each drawn that line in a different place.