🍔 Your Takeaways
There's a real legal AI bubble building-or at minimum, a pricing squeeze that will hit law firm budgets hard in 2026-2027
Three scenarios could play out: price increases (most likely), provider wobbles, or regulatory crackdowns-but they all require the same preparation
Today we're focusing on Scenario 1: the slow price squeeze, where your AI tools still work but cost 30-50% more every renewal
You'll walk away with concrete ways to reduce legal AI costs and lower vendor lock-in risk without abandoning tools that already deliver ROI
Last Monday, Sam Altman sent an internal memo to OpenAI staff declaring a "code red."
Not because of a security breach. Not because of safety concerns. Because Google's Gemini is catching up fast, and OpenAI, despite 800 million ChatGPT users - still isn't profitable.
The company that kicked off the AI boom in 2022 is now scrambling to improve ChatGPT while delaying other products, including AI agents and advertising initiatives.
Meanwhile, HSBC analysts project OpenAI won't reach profitability until well past 2030, with a $207 billion funding shortfall to fill.
For law firms, this isn't just Silicon Valley drama. It's a preview of what happens when AI vendors chase profitability under investor pressure.
Picture this: you open an email from your AI vendor. "We're excited to announce enhancements to our platform..." You scroll down. There it is, a 40% price increase on the tool your partners now rely on daily.
You've got three months to figure out how to absorb that cost, justify it to your CFO, or find an alternative without disrupting workflows. And you know the other vendors are watching to see if anyone pushes back.
I've been thinking a lot lately about what happens to your AI stack if the legal AI bubble doesn't explode dramatically, but quietly deflates.
Not a crash where vendors vanish overnight-something more boring and painful: a gradual pricing squeeze as these companies chase profitability and the cost of infrastructure keeps climbing.
There are really three ways this could go.
Scenario 1 is the slow price squeeze-tools keep working, but costs climb 30–50% at renewal as vendors tighten margins.
Scenario 2 is a major provider wobble-a big AI vendor stumbles through funding issues or gets acquired, and your carefully built integrations break.
Scenario 3 is regulation biting down hard, making everything more expensive and rigid as governments treat AI like critical infrastructure.
All three are plausible. If you forced me to put numbers on it, I’d put Scenario 1 as the most likely-somewhere around 60%-with the other two splitting the rest.
But we’re going to focus on Scenario 1 today because it’s both the most likely and the most actionable.
You can actually do something about it this quarter. You’re not behind. You’re not alone.
And there is a clear, practical path forward that doesn’t require you to rip out systems that are already working.
THE AI BUBBLE
🫧 Three AI Futures for Legal – And Why We're Planning for the Slow Squeeze
Investors are pouring tens of billions of dollars into AI infrastructure, and analysts are openly talking about bubble‑like valuations in AI and legal tech. Some legal AI and infrastructure providers are trading at very high earnings multiples, and commentators have started asking how long that kind of pricing can last.
For your law firm or in‑house team, that pressure translates into three broad scenarios.
Scenario 1: The Slow Price Squeeze (Today’s Focus)
Vendors need to show a path to profitability. The easiest levers are:
Raising per‑seat subscription fees.
Tightening volume‑based pricing (per document, per matter, per query).
Moving customers onto more expensive model tiers “for quality.”
If you’re on an enterprise legal AI platform like Harvey, Lexis+ AI, CoCounsel, or Clio AI, you’ve got lock‑in via contracts, integrations, and user habits.
Proving ROI was hard enough the first time. Now you have to justify a 30–50% increase or tell partners they’re losing a tool they depend on.
If you’re using OpenAI or Anthropic directly, those token invoices that were “background noise” six months ago are starting to attract CFO attention.
OpenAI’s own pricing shows a big spread between tiers: mini‑class models are often in the 10–30% cost band of GPT‑4o‑class per token, while reasoning‑focused models can run roughly 2–4 times more than the baseline.
That spread is the cost lever you should be pulling, but most firms don’t even know it exists.
Scenario 2: A Major Provider Wobble
A flagship AI vendor hits funding trouble, suffers prolonged outages, or gets acquired by a company with different priorities. Your integrations break. Your roadmap evaporates. Your team scrambles to rebuild workflows they thought were stable.
The legal tech market is consolidating quickly, and many AI startups are still burning cash at a rapid rate. It’s not crazy to think one of the names you rely on could change direction.
Scenario 3: Regulation Bites
Governments decide advanced AI needs the same scrutiny as financial services or healthcare. Compliance costs spike. Audit requirements multiply. Flexibility shrinks.
Every workflow needs documented governance, and vendors pass a chunk of those costs straight through to you.
The good news is that preparing for Scenario 1 accidentally makes you safer in Scenarios 2 and 3 as well.
Less vendor lock‑in and better documentation naturally protect you if a provider stumbles.
Solid AI governance-logging, policies, clear workflows-puts you ahead of regulatory pressure whether or not the bubble pops.
Scenario 1 is the one already playing out in slow motion in pricing grids and renewal emails. So let’s focus on that.
PRICE HIKES
💲 How to Survive the Legal AI Bubble's Slow Squeeze
Let me show you three concrete moves that reduce legal AI costs and dependency risk without killing tools that already work.
Part 1: Stop Paying the Fancy Model Tax
Most AI tools quietly default to the newest, most expensive model for everything. It’s like assigning every task to your most senior partner, even routine emails.
Here’s the mental model: use premium models like senior counsel, not for every routine task.
Based on current public pricing, mini‑tier models are often only 10–30% of GPT‑4o‑class cost per token, while reasoning‑focused models can easily run 2–4 times higher.
The exact numbers move, but the spread doesn’t. That’s a potential 10‑plus‑fold cost difference depending on which model you choose for a given task.
The simple guideline: use cheaper mini or previous‑generation models for everyday summarization, drafting, and classification. Reserve high‑end reasoning models for complex, high‑stakes matters where nuance and accuracy justify the cost.
One practical rule you can share internally: “If a junior associate could handle it in under two hours, try a cheaper model first.”
This isn’t about cutting corners. It’s about matching the right tool to the right task, just like you assign work to juniors, mid‑levels, or partners based on complexity.
This is one of the first changes we make when we review AI spend for law firm clients. The savings are immediate and the quality difference for routine tasks is usually negligible.
✂️ Part 2: Choose Cheaper (and Better-Fit) Models For Each Job
Think of this as building a model menu where different models handle different jobs, just like different lawyers handle different matters.
The table below shows you what’s available, what each model does best, and how much cheaper (or more expensive) each option is compared to a GPT‑4o‑class baseline that many firms think of as “standard.”
Model Options to Beat the Fancy Model Tax
All relative costs are approximate, based on current public pricing and typical infrastructure costs. Exact numbers will move, but the shape of the spread is what matters.
Provider / Model | Type | Best Use Case in a Law Firm | Notable Strength | Relative Cost vs GPT‑4o‑class* |
|---|---|---|---|---|
OpenAI – GPT‑4o / GPT‑4.1 | Closed, general flagship | Default “smart” assistant: research memos, complex drafting, analysis | Strong all‑round performance and ecosystem; de facto baseline in many legal AI tools | 1.0x (baseline) |
OpenAI – GPT‑4.1 / 4o mini tiers | Closed, “mini” general LLM | Everyday summarization, email‑level drafting, internal notes | Good quality for routine work at a fraction of GPT‑4o’s price | ~0.1–0.3x |
OpenAI – o1 / o‑series reasoning | Closed, premium reasoning | Hard problems: tricky regulatory analysis, complex strategy docs | Enhanced reasoning; significantly more expensive than GPT‑4o‑class | ~2–4x |
Anthropic – Claude 3.5 / 4.5 Sonnet | Closed, high‑end balanced | High‑stakes drafting where narrative quality matters | Strong competitor to GPT‑4o; similar performance with somewhat higher per‑token costs | ~1.2–1.8x |
Anthropic – Claude 3 / 3.5 Haiku | Closed, fast & cheap “mini” | Bulk summarization, doc triage, quick contract/email Q&A | Very fast and cheap; Anthropic’s budget option for high‑volume workloads | ~0.05–0.2x |
Google – Gemini 1.5 Pro | Closed, long‑context premium | Discovery, due diligence, long transcripts and bundles | Handles up to ~1M tokens; ideal when you need single‑pass analysis of huge document sets | ~0.5–1.0x |
Google – Gemini 1.5 Flash / 3 Flash | Closed, long‑context “mini” | Cheaper large‑document sweeps, fast review of big PDFs/email sets | Lower cost and higher speed than Pro; still supports very long context windows | ~0.2–0.5x |
Llama 3.x / 3.3 70B | Open source, general LLM | Internal legal assistant: drafting, basic analysis, Q&A on firm docs | Strong open‑weight model; fine‑tunable and can run on your infra with no per‑token SaaS fee | Often <0.2x (infra only) |
DeepSeek‑R1 (legal‑leaning) | Open source, reasoning‑heavy | First‑pass contract review, clause spotting, risk flagging | Selected as a top legal open‑source model for its reasoning on legal benchmarks | Often <0.2x (infra only) |
Qwen3‑235B‑A22B (long / multilingual) | Open source, long‑context LLM | Long contracts, multi‑lingual matters, cross‑border work | Multilingual and long‑context; highlighted as a top legal‑industry OSS model | Often <0.2x (infra only) |
Mistral (Small / Large / MoE) | Open source, efficient LLMs | Fast, cheap automations: classification, routing, template drafting | Very good performance‑per‑dollar; runs on modest hardware | Often <0.1–0.2x |
TLDR; from the table above for Associates & Partners.
If you’re using GPT‑4o as your default for everything, you’re leaving money on the table.
For everyday email‑level drafting and summarization, you can move a lot of volume to GPT‑4o‑mini or 4.1‑mini at roughly 10–30% of the cost.
For discovery and due diligence on massive document sets, Gemini 1.5 Pro at roughly half to equal baseline cost gives you a much better long‑context experience.
For truly complex regulatory analysis where you need the best reasoning, you budget for o1‑series models at 2–4x the cost-but only use them where it matters.
For open‑source legal AI models, the savings can be even more dramatic. Llama 3.3 70B, DeepSeek‑R1, Qwen variants, and Mistral models can all run at well under 20% of GPT‑4o’s effective cost because you’re mostly paying for infrastructure, not per‑token SaaS fees.
In some legal benchmarks, these models are now approaching proprietary performance on specific tasks, especially when tuned on legal data. “Cheaper” doesn’t mean “worse”-it means matching the right model to the right job and understanding the relative cost structure in this table.
In practical terms:
- Use Claude Haiku for bulk contract triage.
- Use Gemini Flash for fast sweeps of large discovery bundles.
- Use GPT‑4o‑mini for routine summarization.
- Reserve GPT‑4o or Claude Sonnet for complex drafting.
- And only deploy o1‑series models when you genuinely need premium reasoning on high‑stakes work.
When we build AI automations for clients, we almost always mix models this way.
One for scanning big documents, one for drafting, one for quick classification.
No single model is the cheapest and best at everything, so there’s no reason to pay premium rates across the board.
BUILD CONTINGENCIES
⚠️ Part 3: Take Back Workflow Control (Without Burning Everything Down)
This part is about building optionality so you’re not trapped when renewal comes around. It’s also where the solution gets more optimistic.
For Enterprise Platform Users (Harvey, Lexis+ AI, CoCounsel, Clio AI)
The problem is AI vendor lock‑in for law firms. Contracts, integrations, and user habits make it painful to switch providers, even when prices climb or service quality drops.
But you don’t have to rip out the platform. You just need a backup generator.
Here’s the pattern:
Identify one or two critical workflows where you’re most exposed. Maybe it’s standard NDA review, routine contract first‑pass review, or discovery document triage.
Rebuild simplified versions of those workflows outside the vendor, using your document management system or practice management system plus cheaper models from the menu we just walked through.
Keep the enterprise platform in place. Keep using it. But treat your own workflows as the exit ramp you can scale if prices jump or the vendor stumbles.
This doesn’t mean becoming a dev shop. One internal champion plus an external developer or consultant can stand up a few key automations in weeks, not months.
This is a pattern we’ve used when clients want to de‑risk a Harvey‑style rollout: keep what works today, but quietly build an exit ramp in the background so renewal negotiations happen on your terms, not theirs.
For Direct OpenAI or Anthropic Users: Build a Knowledge Cache and a Backup Path
If you’re hitting OpenAI or Anthropic directly, your main risk in Scenario 1 is creeping spend and over‑reliance on one provider’s interface.
Start by building a knowledge cache:
Use features like Claude’s Projects to collect source docs, prompts, and outputs for a workflow in one place, so your team isn’t reinventing the wheel every time. When the project’s knowledge approaches context limits, those tools can expand capacity while maintaining response quality.
Mirror that cache inside your own systems-Notion, Confluence, or your DMS. Create “best prompts and patterns” pages per workflow. Treat them like internal playbooks, not personal chat logs.
The payoff is simple: the more you standardize how you use AI, the easier it is to shift that usage between models or providers when costs change. You’re no longer paying over and over for the same thinking; you’re compounding what you’ve already learned.
If the legal AI bubble does start to deflate and prices rise, you won’t be scrambling. You’ll already know which workflows to shift to cheaper models, and you’ll have the playbooks and patterns documented so switching is a decision, not a fire drill.
That’s how law firms can prepare for an AI bubble without panicking or overspending.
🗞️ Related Legal AI News
Borrowing by AI companies represents a 'mounting potential threat to the financial system,' top economist says – Big Tech has accumulated nearly $250 billion in debt this year alone to fund AI infrastructure, raising systemic risk concerns
The question isn't whether the AI bubble will burst – but what the fallout will be – Financial analysts examine what the AI bubble will leave of value once it deflates and how the fallout will impact the broader economy.
🛠️ 10 Second Explainers - AI Tools & Tech
AI vendor lock-in for law firms When contracts, integrations, and user habits make it prohibitively expensive to switch providers, even when prices climb or service quality drops. Preparing for Scenario 1 means building optionality so you're not trapped at renewal time.
Model abstraction layer Think of it as a universal plug adapter that lets you swap AI models without rebuilding your entire workflow. You change the model "behind the scenes" while the user experience stays the same. It's what keeps you flexible when prices or providers change.
Knowledge cache A shared, reusable library of prompts, outputs, and patterns that makes your AI usage teachable and portable. Instead of every lawyer starting from scratch, your best prompts and workflows live in one place and can move between providers if needed.
"I think no company is going to be immune, including us. I expect AI to be the same [as the dot-com era]."
READER POLL
If your main AI tools raised prices by 40% next year, what would you do first to reduce legal AI costs?
A) Cut seats and limit access to essential users only
B) Negotiate hard and threaten to walk
C) Shift workflows to cheaper models and open-source alternatives
D) Pause AI expansion until the market stabilizes
[Reply with your letter choice] - I'll share the results in the next edition.
My Final Take…
You don’t need to predict winners and losers in the AI market. That’s not your job.
What you do need is a stack and cost structure that survives a legal AI bubble, whether it pops dramatically or deflates slowly over the next two years.
Preparing for Scenario 1-the gradual price squeeze-is the most practical way to get there. Stop paying the fancy model tax. Build a model menu that matches tools to tasks. And create optionality so renewal conversations happen on your terms, not theirs.
That’s the path forward, and it’s one you can start walking this quarter.
- Liam Barnes
If you’d like a more tailored review-a light‑touch “AI cost and dependency review” or a “Scenario‑1 stress test” for your specific stack.
Grab some time to chat
(if you don’t see a suitable time, just shoot me an email [email protected])
How Did We Do?
Your feedback shapes what comes next.
Let us know if this edition hit the mark or missed.
Too vague? Too detailed? Too long? Too Short? Too pink?
Was this week’s newsletter forwarded to you?
Sign up, it’s free.
Last Week’s Reader Poll Results
