Agent-platform pricing is hard to predict — and what a 'good' model looks like

Ask "what will this agent cost me next month?" and most platforms can't give you a straight answer. That is why our pricing score measures predictability and transparency, not headline cheapness. A platform you can forecast scores higher than one that's technically inexpensive but billed across a dozen moving parts.

The metering maze

Across the platforms we track, the number of distinct billing units is remarkable:

• Credits — Lindy, Dify, Replit Agent (effort/credits), and others.
• Activities — Zapier Agents counts a trigger, knowledge answer, action run, web browse, or search each as one activity.
• Actions + Vendor Credits — Relevance AI bills on two axes at once.
• Traces, runs, uptime, LCUs — LangGraph's managed tier meters across several dimensions.
• Per-second compute — Modal, Fly.io Machines, and E2B bill by the second, sometimes atop a monthly base.

Each model is defensible in isolation. Stacked across a shortlist, they make apples-to-apples comparison genuinely hard — which is the point of our per-platform pricing notes on /platforms.

Where transparency breaks down

Two failure modes recur:

• Multi-axis metering. When cost depends on three or four independent counters, you can't build a simple forecast. Relevance AI's Action-plus-Vendor-Credit split is the clearest example.
• Undisclosed figures. Some platforms describe the model but not the numbers. We flag Alfe for exactly this — a credit-pool-plus-subscription model that's described publicly but without a verifiable tier table — and Zapier Agents' paid-tier figures were unavailable at access time.

What a good pricing model looks like

The highest pricing scores in our catalog go to the SDKs — the Vercel AI SDK and OpenAI Agents SDK — both free and open source (MIT), where you simply pay the underlying model's token costs. That's the cleanest possible story: one variable you already understand.

Short of that, a good model tends to share these traits:

• Few metered dimensions, ideally one or two.
• Published numbers you can plug into a spreadsheet.
• A free tier or open-source core to test on — see free to start.
• A predictable base, so a busy month doesn't produce a surprise invoice.

The practical advice: before committing, model a realistic month against the actual metered units, not the headline price — and get any undisclosed figures in writing.

Key takeaways

• We score pricing on predictability, not cheapness.
• Credits, activities, actions, traces, runs, and per-second compute make cross-platform comparison hard.
• The cleanest models are the open SDKs (Vercel AI SDK, OpenAI Agents SDK): free, you pay only token costs.
• Watch for multi-axis metering and undisclosed figures; model a real month before you commit.