---
title: "AI Cost Math for Law Firms: Per-Seat vs Usage-Based in 2026"
slug: "ai-cost-math-for-law-firms-per-seat-vs-usage"
author: "ibl.ai Engineering"
date: "2026-05-30 11:00:00"
category: "Premium"
topics: "AI cost legal, law firm AI pricing, Harvey AI pricing, CoCounsel pricing, legal AI per-seat, contract review AI, due diligence AI, GPT-5 legal, Claude Opus legal, self-hosted legal AI, privilege protection AI"
summary: "What AI actually costs an AmLaw firm in 2026 — token pricing for the latest models against the $300–500/lawyer/month Harvey and Co:Counsel bills, with the privilege math for contract review and due diligence at scale."
banner: ""
thumbnail: ""
---

## The AmLaw 100 Math: $400 × 200 Lawyers Is Not the Right Number

Harvey lists around $300–500 per lawyer per month. Thomson Reuters' Co:Counsel runs $200–500. A 200-lawyer firm staring at a $400/seat bill is looking at **$80,000 per month — close to $1M per year** — for AI that almost certainly doesn't get used by every seat every day.

The pricing model is borrowed from legal-research databases (Westlaw, LexisNexis) — where every lawyer needs occasional access and the per-seat fee is a rounding error on a $1M+ partner. For AI doing the actual work — first-pass contract review, due-diligence summaries, deposition prep — the seat assumption breaks. Most lawyers touch it once a week; a few use it constantly.

Token pricing aligns the bill to the work. Self-hosting puts the privileged data behind your own firewall. The math is the post.

## What the Latest Models Actually Cost in 2026

Token pricing across the major providers, approximate as of mid-2026:

<table style="width:100%; border-collapse:collapse; margin:1.5rem 0; font-size:0.95rem;">
  <thead>
    <tr style="background:#f5f5f0; border-bottom:2px solid #2175C5;">
      <th style="text-align:left; padding:0.75rem; color:#5f6368;">Model</th>
      <th style="text-align:left; padding:0.75rem; color:#5f6368;">Provider</th>
      <th style="text-align:right; padding:0.75rem; color:#5f6368;">Input ($/MTok)</th>
      <th style="text-align:right; padding:0.75rem; color:#5f6368;">Output ($/MTok)</th>
      <th style="text-align:left; padding:0.75rem; color:#5f6368;">Best for</th>
    </tr>
  </thead>
  <tbody>
    <tr style="border-bottom:1px solid #e5e7eb;">
      <td style="padding:0.75rem;"><strong>Claude Opus 4.7</strong></td>
      <td style="padding:0.75rem;">Anthropic</td>
      <td style="text-align:right; padding:0.75rem; font-variant-numeric:tabular-nums;">$15</td>
      <td style="text-align:right; padding:0.75rem; font-variant-numeric:tabular-nums;">$75</td>
      <td style="padding:0.75rem;">Complex contract analysis, brief-writing</td>
    </tr>
    <tr style="border-bottom:1px solid #e5e7eb;">
      <td style="padding:0.75rem;"><strong>Claude Sonnet 4.6</strong></td>
      <td style="padding:0.75rem;">Anthropic</td>
      <td style="text-align:right; padding:0.75rem; font-variant-numeric:tabular-nums;">$3</td>
      <td style="text-align:right; padding:0.75rem; font-variant-numeric:tabular-nums;">$15</td>
      <td style="padding:0.75rem;">First-pass review, summarization</td>
    </tr>
    <tr style="border-bottom:1px solid #e5e7eb;">
      <td style="padding:0.75rem;"><strong>Claude Haiku 4.5</strong></td>
      <td style="padding:0.75rem;">Anthropic</td>
      <td style="text-align:right; padding:0.75rem; font-variant-numeric:tabular-nums;">$1</td>
      <td style="text-align:right; padding:0.75rem; font-variant-numeric:tabular-nums;">$5</td>
      <td style="padding:0.75rem;">Document classification, intake triage</td>
    </tr>
    <tr style="border-bottom:1px solid #e5e7eb;">
      <td style="padding:0.75rem;"><strong>GPT-5</strong></td>
      <td style="padding:0.75rem;">OpenAI</td>
      <td style="text-align:right; padding:0.75rem; font-variant-numeric:tabular-nums;">$10</td>
      <td style="text-align:right; padding:0.75rem; font-variant-numeric:tabular-nums;">$30</td>
      <td style="padding:0.75rem;">Deposition prep, legal research</td>
    </tr>
    <tr style="border-bottom:1px solid #e5e7eb;">
      <td style="padding:0.75rem;"><strong>Gemini 3 Pro</strong></td>
      <td style="padding:0.75rem;">Google</td>
      <td style="text-align:right; padding:0.75rem; font-variant-numeric:tabular-nums;">$3.50</td>
      <td style="text-align:right; padding:0.75rem; font-variant-numeric:tabular-nums;">$10.50</td>
      <td style="padding:0.75rem;">Long-context review (multi-MM token files)</td>
    </tr>
    <tr style="border-bottom:1px solid #e5e7eb;">
      <td style="padding:0.75rem;"><strong>Llama 4 (70B, self-hosted)</strong></td>
      <td style="padding:0.75rem;">Meta (open weights)</td>
      <td style="text-align:right; padding:0.75rem; font-variant-numeric:tabular-nums;">~$0</td>
      <td style="text-align:right; padding:0.75rem; font-variant-numeric:tabular-nums;">~$0</td>
      <td style="padding:0.75rem;">Privileged work behind firewall</td>
    </tr>
    <tr style="border-bottom:1px solid #e5e7eb;">
      <td style="padding:0.75rem;"><strong>DeepSeek-R1 (self-hosted)</strong></td>
      <td style="padding:0.75rem;">DeepSeek (open weights)</td>
      <td style="text-align:right; padding:0.75rem; font-variant-numeric:tabular-nums;">~$0</td>
      <td style="text-align:right; padding:0.75rem; font-variant-numeric:tabular-nums;">~$0</td>
      <td style="padding:0.75rem;">Cost-sensitive bulk review</td>
    </tr>
  </tbody>
</table>

For self-hosted open-weight models, the marginal cost is GPU time. A reserved H100 instance ($1.50–3/hour) handles tens of thousands of contract pages per day.

## A Real Workload: Due-Diligence Contract Review at a 200-Lawyer Firm

Take a mid-market M&A practice running first-pass contract review on a typical deal: **5,000 documents per deal, 6 deals per month = 30,000 contracts per month**. Each contract is roughly 3,000 input tokens (the document) and 800 output tokens (the structured summary with flagged clauses and risk classifications). For a deeper per-contract cost breakdown — including a side-by-side against Harvey, Co:Counsel, Spellbook, Ironclad AI, and LinkSquares at three scale tiers (solo / mid-market / AmLaw) — see **[What AI Contract Review Actually Costs in 2026](/blog/what-ai-contract-review-actually-costs-2026)**.

That's **90M input + 24M output tokens per month** across the entire firm — front-loaded onto the M&A and corporate teams that actually run diligence.

### What it costs by deployment shape

<table style="width:100%; border-collapse:collapse; margin:1.5rem 0; font-size:0.95rem;">
  <thead>
    <tr style="background:#f5f5f0; border-bottom:2px solid #2175C5;">
      <th style="text-align:left; padding:0.75rem; color:#5f6368;">Deployment</th>
      <th style="text-align:left; padding:0.75rem; color:#5f6368;">Pricing shape</th>
      <th style="text-align:right; padding:0.75rem; color:#5f6368;">Monthly cost</th>
      <th style="text-align:right; padding:0.75rem; color:#5f6368;">Annual</th>
      <th style="text-align:left; padding:0.75rem; color:#5f6368;">Privilege posture</th>
    </tr>
  </thead>
  <tbody>
    <tr style="border-bottom:1px solid #e5e7eb;">
      <td style="padding:0.75rem;"><strong>Harvey</strong></td>
      <td style="padding:0.75rem;">Per-lawyer (~$400)</td>
      <td style="text-align:right; padding:0.75rem; font-variant-numeric:tabular-nums; color:#b91c1c;"><strong>$80,000</strong></td>
      <td style="text-align:right; padding:0.75rem; font-variant-numeric:tabular-nums; color:#b91c1c;">$960,000</td>
      <td style="padding:0.75rem;">Harvey-hosted (their cloud, their model choice)</td>
    </tr>
    <tr style="border-bottom:1px solid #e5e7eb;">
      <td style="padding:0.75rem;"><strong>Thomson Reuters Co:Counsel</strong></td>
      <td style="padding:0.75rem;">Per-lawyer (~$300)</td>
      <td style="text-align:right; padding:0.75rem; font-variant-numeric:tabular-nums; color:#b91c1c;"><strong>$60,000</strong></td>
      <td style="text-align:right; padding:0.75rem; font-variant-numeric:tabular-nums; color:#b91c1c;">$720,000</td>
      <td style="padding:0.75rem;">TR-hosted (their cloud, their model choice)</td>
    </tr>
    <tr style="border-bottom:1px solid #e5e7eb;">
      <td style="padding:0.75rem;"><strong>ChatGPT Enterprise</strong></td>
      <td style="padding:0.75rem;">Per-seat ($60/user, ~250 staff)</td>
      <td style="text-align:right; padding:0.75rem; font-variant-numeric:tabular-nums; color:#b91c1c;"><strong>$15,000</strong></td>
      <td style="text-align:right; padding:0.75rem; font-variant-numeric:tabular-nums; color:#b91c1c;">$180,000</td>
      <td style="padding:0.75rem;">OpenAI cloud (general-purpose, not legal-tuned)</td>
    </tr>
    <tr style="border-bottom:1px solid #e5e7eb;">
      <td style="padding:0.75rem;">Direct API — Claude Sonnet 4.6</td>
      <td style="padding:0.75rem;">Token-based</td>
      <td style="text-align:right; padding:0.75rem; font-variant-numeric:tabular-nums;">~$630</td>
      <td style="text-align:right; padding:0.75rem; font-variant-numeric:tabular-nums;">~$7,560</td>
      <td style="padding:0.75rem;">Anthropic cloud (firm DPA)</td>
    </tr>
    <tr style="border-bottom:1px solid #e5e7eb;">
      <td style="padding:0.75rem;">Direct API — GPT-5</td>
      <td style="padding:0.75rem;">Token-based</td>
      <td style="text-align:right; padding:0.75rem; font-variant-numeric:tabular-nums;">~$1,620</td>
      <td style="text-align:right; padding:0.75rem; font-variant-numeric:tabular-nums;">~$19,440</td>
      <td style="padding:0.75rem;">OpenAI cloud (firm DPA)</td>
    </tr>
    <tr style="background:#f0f9ff; border-bottom:1px solid #e5e7eb;">
      <td style="padding:0.75rem;"><strong>ibl.ai self-hosted (Llama 4 / DeepSeek-R1)</strong></td>
      <td style="padding:0.75rem;">Flat license + GPU</td>
      <td style="text-align:right; padding:0.75rem; font-variant-numeric:tabular-nums; color:#15803d;"><strong>~$5,000–8,000</strong></td>
      <td style="text-align:right; padding:0.75rem; font-variant-numeric:tabular-nums; color:#15803d;">~$60,000–96,000</td>
      <td style="padding:0.75rem;"><strong>Inside the firm's network (privilege intact)</strong></td>
    </tr>
  </tbody>
</table>

The ibl.ai row covers the GPU instance, the platform license, and ongoing support. There is no third-party vendor in the data path, no managed-cloud DPA to renegotiate, and no question about whether the model provider could be served with a subpoena that reaches your privileged work product.

## Why Per-Seat Pricing Fails Harder in Legal

Three reasons the seat model breaks for law firms specifically:

**1. Usage is bimodal.** M&A, corporate, and litigation eDiscovery teams hit AI constantly. Real estate, T&E, and most non-transactional groups touch it occasionally. Buying the same seat for both means the heavy users subsidize the firm's average — but the firm pays the seat fee for everyone.

**2. The "lawyer count" you're billed on includes a lot of non-billable seats.** Paralegals, document-review attorneys, and staff are often counted at the same per-seat rate even though their billing rate is a fraction of a partner's. The vendor's "per-lawyer" pricing is really "per-headcount."

**3. Privilege is structurally incompatible with managed clouds.** ABA Model Rule 1.6 obligates lawyers to make "reasonable efforts to prevent the inadvertent or unauthorized disclosure of" client information. Sending privileged documents to a third-party AI vendor — even one with a DPA and SOC 2 — introduces a custody question that doesn't exist when the model runs on the firm's infrastructure. Several jurisdictions are now writing this into formal opinions.

## What Stays the Same, What Changes

Self-hosting the runtime doesn't mean rebuilding the firm's tooling. The chat UI, the matter-scoped workspaces, the citation-checking, the document-management integration (iManage, NetDocuments, SharePoint), the multi-agent orchestration — all of that stays managed by ibl.ai. The compute, the model, and the privileged data move inside the firm's network.

What disappears: the $1M/year per-seat line item. What appears: an internal capability the firm owns and controls, with the same agent-orchestration platform and the freedom to pick the model that fits each workload — Opus for the brief, Sonnet for the contract sweep, an open-weight model for the bulk diligence run.

## Run the Numbers for Your Firm

For workload sizing and cost modeling specific to your practice areas, the **[AI Help Desk Cost Savings Calculator](/resources/calculators/ai-help-desk-savings-calculator)** generalizes to most high-volume legal-administrative workloads (intake triage, document classification, first-pass review).

For the deployment comparison side-by-side — including ABA Model Rule 1.6 posture, privilege protection, and air-gapped options for the most sensitive matters — see **[Self-Hosted AI vs ChatGPT Enterprise for Legal](/resources/comparisons/self-hosted-ai-vs-chatgpt-enterprise-for-legal)**.

For the broader policy framework — what a law-firm AI policy should cover and why owned/air-gapped deployment is the control that makes it enforceable — see **[AI Policies for Law Firms: A Practical 2026 Guide](/blog/ai-policies-for-law-firms)**.

## Why Family-Owned and New York Matters Here

A law firm's AI vendor relationship is a long-term commitment — the workflows, the prompts, the integration with the document-management system, the audit logs the firm relies on for malpractice defense. ibl.ai is family-owned and operated from New York, NY — a long-term partner with a perpetual platform license and no investor exit pressure. The runtime is open source. The privileged data stays inside the firm's network. The math works at 50 lawyers or 2,000.