The legal AI landscape, decoded for lawyers.

LLM stands for Large Language Model. It's the technology behind ChatGPT, Claude, and Gemini. These models are trained on hundreds of billions of words scraped from the internet, books, and code. That training lets them predict the next word in a sequence with uncanny accuracy, which turns out to produce surprisingly useful writing, reasoning, and summarization.

They are genuinely powerful. But they were trained to produce plausible text, not accurate law. They have no access to a verified case law database. They don't know what was true as of yesterday. And left unchecked, they confidently fabricate citations that don't exist. This is called hallucination.

Using a raw LLM for legal work is like hiring a paralegal who has read every book in the library but has never verified a single citation and cannot tell you if a case was overruled.

LLMs don't read words. They read tokens. A token is roughly three-quarters of a word. "attorney" is one token; "attorney-client" is two or three. Code and unusual words can cost more. Numbers, punctuation, spaces: all tokens.

Why does this matter? Because everything about AI performance and cost is measured in tokens. The "context window" (how much the model can see at once) is measured in tokens. A 50-page contract is roughly 25,000 to 35,000 tokens. A full deposition transcript might be 60,000 tokens. Some models have 128,000-token windows; others, 200,000.

When you hit the context limit, the model can no longer see the beginning of the document while it reads the end. For legal work, where a defined term on page 3 governs a clause on page 47, that is not a theoretical problem. It is a real one.

Every call to an LLM API costs money per token, charged separately for input (what you send) and output (what you receive). OpenAI and Anthropic publish their rates; as of 2026, frontier models cost between $3 and $75 per million tokens depending on the model.

A single complex legal research query, where you upload a contract and ask for a risk analysis, might consume 40,000 to 100,000 tokens. That is $0.12 to $7.50 in raw API cost per query, before any infrastructure, engineering, or margin.

This is why "unlimited AI" claims from wrapper products deserve scrutiny. Someone is paying for those tokens. Either usage is quietly throttled, or you are, at a steep markup.

A wrapper is a product built by taking an existing LLM API (OpenAI's GPT, Anthropic's Claude), adding a user interface, writing a system prompt that says something like "You are a helpful legal assistant," and selling the result as a legal AI product.

There is no proprietary technology. No legal-specific training. No citation verification infrastructure. No independent data architecture. The wrapper company is a middleman between you and OpenAI's API, typically charging 5 to 10 times the underlying API cost.

The risks are real: your documents flow through OpenAI or Anthropic's servers, which have been subpoenaed. The model has no awareness of attorney-client privilege. Hallucinated citations get returned with the same confidence as real ones. And when the underlying model improves or changes, the wrapper has no control over the behavior of its own product.

Several well-known legal AI tools, including some with significant venture backing and aggressive marketing, are wrappers. You can often identify them because they cannot explain their citation verification methodology, their data never leaves third-party servers, and their pricing increases significantly for features that a purpose-built platform includes by default.

Irys is a purpose-built legal platform. Not a chatbot with a law firm logo, and not a wrapper around someone else's API.

The architecture has three stages. In the first stage, your documents and matter context are processed by Irys's own retrieval and analysis pipelines: text is extracted, structured, indexed, and made queryable. In the second stage, an agentic reasoning layer (built by Irys, not rented from a lab) plans the research task, queries the case law database, validates citations against primary sources, and resolves conflicts. Only in the third stage does a frontier model receive a prompt, and by that point, the prompt is small, precise, and grounded in verified facts.

The result: fewer hallucinations (because the model isn't guessing at case law, it's synthesizing retrieved and verified results), better performance on complex multi-document tasks, genuine privilege protection (data never touches third-party model servers), and pricing that reflects actual infrastructure costs rather than API markup.

The practical difference for you: research that cites real cases, drafts that reflect your actual document set, and a platform that handles the full legal workflow (research, drafting, redlining, matter management, document comparison) in one place.