The unit problem

27 of 47 products publish a unit price. They are counting 14 different things. The cheapest published unit is $0.0012 and the dearest is $30 — a 25,000x spread that means nothing, because a credit is not a token is not a minute is not a resolution.

This page does what a pricing teardown does first: puts every number on one axis, says which ones were read and which ones we computed, keeps comparison to the same meter, and then asks the only question a buyer has — what does one unit of my work cost? — using nothing but figures on the vendor's own page.

Every unit price in the index on one log axis, grouped by what is being counted

Filled marks were read off the page. Hollow volt marks were computed by IY from two published figures (the arithmetic is in section 04). The ice diamond is a unit IY re-expressed so it could sit on the same line as its family. Hover any mark for the vendor's own words.

Where a comparison is legitimate, and the one caveat each row carries

The token row is the only clean benchmark in the set: 7 vendors, one unit, a 23x spread on list — and every one of the big three then discounts reuse so steeply that the list price is the ceiling, not the price. The credit row is the opposite: 8 vendors, one word, and no two of them mean the same thing by it.

What a task, a conversation, a minute or a million tokens can cost — computed from the vendor's page, nothing else

For each vendor whose page gives enough to do the arithmetic, the bar runs from the cheapest case the page allows to the dearest. An open arrow means the page names a further charge without bounding it. The ice tick is the headline on the plan card. Width is the finding: the wider the bar, the less a buyer can forecast, and the more a pricing conversation is worth. Every bar is DERIVED; hover for the arithmetic.

Not computable from the page

15 more products publish a number that cannot be turned into the cost of one piece of work. Each absence is a specific, fixable omission.

• Gumloop$0.005 per credit is published, but a task costs 'tokens, compute and paid tool calls at each provider's list price, plus an 8% orchestration fee' (16% with your own keys) — a pass-through, so the page cannot price a task
• GitHub Copilot1 AI credit = $0.01 is pegged to the dollar; how many credits one request consumes is not on the page
• Apollo.io$0.0196 per credit; credits per task are not stated
• Zapier Agents$0.0222 per activity; an activity is not defined on the pricing page
• Notion AI$0.01 per credit once normalized; credits per agent action are not stated
• Claythe same Launch tier reads $54 or $167 depending on slider position; actions included at each point are not stated
• Cursor'$20/mo' buys an 'included model usage pool, then on-demand'; the pool is not sized on the page
• Replit$20 of model allowance per month — the unit is a dollar, and what a dollar buys depends on the model chosen
• Devin · Windsurfpriced per message with 'extra usage consumed at API pricing'; neither the message allowance nor the per-message rate is on the page
• Claude (Pro / Max / Team)capacity sold as '5x / 20x' multiples of Pro; no unit is named
• Zendesk AI agents'pricing based on the successful outcomes they deliver' with a resolution allowance per plan; no per-resolution dollar figure is published
• Bolt.newtoken-metered with rollover; tokens per plan and per prompt are not on the page
• Perplexity$20 / $200 seats; no usage unit named
• n8n Cloudprices in EUR only ('20€/mo'); no USD figure exists to record
• Synthflow AIa single Enterprise plan 'starting at $30,000 annually'; no unit

Every derived figure on this page, with the two published numbers it came from

A derived unit price never appears unmarked — that is the whole promise of the format. If a figure above disagrees with the index it is flagged, not averaged.