We read 100 AI agent pricing pages. Here's what's broken.

Pricing is one of the clearest signals of how a product works. If a pricing page is confusing, the product often is too. Over the last month we looked at 100 AI agent and AI automation pricing pages — indie SaaS to enterprise platforms — and found four patterns that keep showing up. Three are broken. One is starting to work.

Pattern 1: effort-based pricing (mostly broken)

Pricing varies by 'effort' of the request. Complex thing = more credits. A simple edit might be $0.10; a complex feature $5+. Replit Agent 3 is the most famous example. Lindy has task-tier pricing that's adjacent.

What's broken: unpredictable monthly bills. Users report $100–$300 in top-ups on top of subscription. Budgeting becomes guesswork. Teams that adopted the product eagerly back off when the invoice comes.

What it's actually good for: one-off tasks. A person building a project should pay more for harder work. For recurring work — every day, every workflow — effort-based pricing punishes the best customers.

Pattern 2: per-seat pricing (half-broken)

Classic SaaS carry-over. Claude Team Premium is $100/seat (5-seat minimum). ChatGPT Business $25-30/user. Dust and Glean do per-seat on top of enterprise base.

What's broken: assumes agents map to humans 1:1. But an agent is not a user. A team of five people might run twenty agents. Or one person might run a swarm of fifty. Per-seat pricing either overcharges (few humans, many agents) or undercharges (many humans, few agents).

Where it still works: consumer-adjacent AI where the agent is a glorified chatbot, one-per-user.

Pattern 3: token/API pricing (broken for most buyers)

Raw usage-based: $3–$15 per million input/output tokens. Anthropic and OpenAI APIs. Pretty much every self-hosted/framework-based setup.

What's broken: the buyer has to model usage. Most don't know their token footprint, so they guess low and get surprised. Engineering teams accept this because they run optimizations; non-technical buyers can't.

Where it works: developer tools. Not consumer. Not operations. Not knowledge-work.

Pattern 4: flat credit pool (starting to work)

Monthly fee, fixed credits, every feature included. Spawnlabs plans, ChatGPT Pro, and a handful of newer platforms. User pays a predictable amount; compute is handled internally.

Why it's better: budgeting is simple. Heavy use is rewarded (agent compounds value). Light use isn't punished (overages are rare on well-sized plans).

Where it breaks: if the plan is badly sized for your usage. Underprice and the vendor hemorrhages; overprice and the buyer feels gouged. The operator on each side has to actually do the math, which most don't. When it works, it's the right shape.

The invisible cost: model volatility

Every pricing page has one unspoken clause: 'prices may change with model upgrades.' Translation — if Claude 5 costs 2× Claude 4.6, your monthly invoice goes up. Most vendors pass model costs through. Some absorb it. Nobody is transparent about how much room there is before your bill moves.

What to ask a vendor:

• What's the pricing model — effort, seat, token, credit?
• What counts as a billable unit at each tier?
• What's the typical top-up rate for teams my size?
• How much does pricing shift when models change?
• Is there a cap / overage buffer?

What we do at Spawnlabs

Flat credit pool. $39/mo Starter, $299 Growth, $599 Scale, custom Enterprise. Every agent, every tool the agent builds, every integration — included. Credits are priced so overages are rare for properly-sized plans; when they happen, top-ups are at published rates, not surprise.

We think effort-based is the worst fit for recurring work, per-seat is a holdover from old SaaS, and token-based belongs in developer tools. Flat credits isn't perfect, but it's the model we'd buy if we were buyers.

The pricing page reveals the product. Now you know how to read it.