I Read 25 AI Pricing Pages. A 'Credit' Is $1 at JetBrains and One Cent at Perplexity.

TL;DR: One AI credit buys whatever the vendor decided it buys this quarter, and there is no cross-vendor answer. On August 15, 2026 I read the pricing page and the linked billing docs for 25 AI tools and counted 18 distinct names for the billing unit. Eleven of the 25 call it a "credit," and the published rates put a JetBrains AI credit at $1.00 and a Perplexity Computer credit at $0.01, a hundredfold spread on the same word. Four vendors state no rate anywhere public, and five express included usage only as a multiple of a base number they never print.

Seats were comparable: two vendors at $30 a seat sold the same quantity of the same thing. Credits are not, and that is what they are for.

What did I actually count?

One method, applied 25 times: open the pricing page, follow every billing link, and record four things. What the unit is called. Whether the page puts a rate on it. Whether a rate exists anywhere public. How many clicks to reach it. Two pages needed a real browser to render: JetBrains and Devin.

Two of the 25 no longer have a pricing page. continue.dev/pricing returns 404 because Continue was acquired by Cursor. And codeium.com/pricing 301s to windsurf.com/pricing, which 308s to devin.ai/pricing. Two renames and an acquisition, and what was Codeium's pricing page now lands on a different product line.

Tool	Unit it bills in	On the page?	Anywhere public?	Clicks
Cursor	"3x / 20x Pro limits"	No	Partly, in dollars	1
GitHub Copilot	premium request	No	Yes, plus multipliers	1
Claude Code	"5x or 20x more usage"	No	No	n/a
Zapier	task, plus Agent "activity"	Yes	Yes	0
n8n	workflow execution	Yes	Yes	0
Make	credit (formerly operation)	Yes	Yes	0
Replit	credit, plus checkpoint	Yes, in dollars	Yes	0
Lovable	credit	Partly, in Plan Mode	Partly	0
Devin	"quota"; ACU in docs	No	No rate	n/a
Windsurf	none, redirects to Devin	n/a	n/a	n/a
OpenAI Codex	credit, message, task	No	Yes, per million tokens	2
Gemini Code Assist	requests per day	No	Yes: 6,000 code, 960 chat	2
v0	credit, plus token	Yes, in dollars	Yes	0
Bolt	token	Yes: 10M/month on Pro	Yes	0
Warp	credit (AI, compute, platform)	Yes: 1,500 = $20	Yes	1
Cline	AI inference credit	No	Yes, pass-through	1
Sourcegraph Amp	dollars of usage, orb hours	Yes, in dollars	Yes	1
Factory	"Standard Usage", ~5x / ~10x	No	No	n/a
Tabnine	seat	Yes, per seat	Yes	0
Qodo	credit	Yes: $0.012 each	Yes	0
Continue	none, page 404s	n/a	n/a	n/a
Augment	dollars of usage	Yes, list plus 40%	Yes	0
JetBrains AI	AI credit	Page needs JavaScript	Yes: 1 credit = $1	1
Zed	token	Yes	Yes	0
Perplexity	Computer credit	No, counts only	Yes: 100 credits = $1	3

The 18 unit names, so you can check the count: credit, premium request, task, activity, workflow execution, operation, checkpoint, ACU, request, token, dollar of usage, orb hour, seat, accepted edit prediction, message, review, quota, and Factory's unnamed "Standard Usage."

What is one credit actually worth?

Four vendors publish a credit-to-dollar rate and they disagree by two orders of magnitude. JetBrains: "Each AI Credit equals $1 USD," with 3, 10 and 35 per 30 days on Free, Pro and Ultimate. Perplexity, updated August 6, 2026: "Today, 100 credits equals $1," with Max including 10,000 a month and a light task costing 100 to 350. Qodo prints "$.012/credit" on its pricing page. Warp sells "1,500 credits ($20 of included agent usage at API rates)."

So: $1.00, $0.01, $0.012, roughly $0.013. Same word, four currencies, no exchange rate. Make's credit is not a currency at all ("Each module action in your scenario, like adding a Google Sheet row, counts as one credit"). Lovable's is a message, sometimes. OpenAI's docs say GPT-5.6 "averages 5-40 credits per message," making one credit a fraction of a prompt.

Which vendors never say?

Four, and not small ones. Anthropic offers "5x or 20x more usage than Pro," and its support article on usage limits explains only that usage "is affected by several factors, including the length and complexity of your conversations." Cursor sells "Extended limits," "3x Pro limits" and "20x Pro limits" on Agent and quantifies none of them (its docs price a separate model pool in dollars and call the Agent pool "generous"). Factory sells Plus at "~5x Pro usage" with no baseline anywhere. Devin deleted its unit outright.

That construction appears on five of the 24 live pages: Cursor, Anthropic, OpenAI, Factory, and Perplexity, whose table lists "10x" against a Free column reading "Limited." A multiple of an undisclosed base is not a quantity. It is the shape of one, and it is the fog that makes a Cursor versus Claude Code comparison unpriceable.

Did any vendor publish a number and then stop?

Three, all in the summer of 2025.

Devin. The Wayback Machine's June 16, 2025 capture lists a $500 a month Team plan with "250 ACUs included monthly," or $2 per Agent Compute Unit. Today the live page reads "Light quota," "Increased quotas," and "Significantly higher quotas." The letters A-C-U appear nowhere on it.

Replit. Its June 18, 2025 announcement is candid: "Previously, the Agent charged $0.25 per checkpoint." The replacement is effort-based, where "simple tasks may cost less than $0.25, more complex tasks may cost more than $0.25." A fixed price became a range with no stated ceiling.

Make. It renamed operations to credits at 1:1, and its help center draws the distinction precisely: "Operations count the activities performed in your scenarios, while credits are what you pay for those activities." Non-AI apps still run 1 operation to 1 credit. AI features can consume more. The rename is what let the two drift apart.

Is metering by usage actually wrong?

No, and the counter-examples are why this is about disclosure, not metering. Inference costs a variable amount per request, and flat per-seat pricing either overcharges light users or bankrupts you on heavy ones. GitHub solved that in public: its billing docs define a premium request as "any interaction where you ask Copilot to do something for you," give 300 a month on Pro and 1,500 on Pro+, price overage at $0.04, and publish dated per-feature multipliers including this one: "Starting June 1, 2026, Copilot code review will have a model multiplier of 13." Unflattering, and printed anyway.

Sourcegraph's Amp passes inference through at zero markup, with a worked example: a thread incurring $2 of Anthropic and $0.50 of OpenAI deducts $2.50. Zed defines a token on the pricing page ("roughly 3-4 characters of English text") and bills at list plus 10%. None of them sell a fixed unit. All of them are priceable.

What is the one question to ask?

Ask a vendor: how many of your units does one dollar buy, and where is that written down? JetBrains, Perplexity, Qodo, Warp, Amp, Zed and Augment answer in one sentence with a URL. A vendor selling "3x limits" cannot answer at all, and not because the answer is complicated. A published rate is comparable, and a comparable price is one you can shop. The support-agent vendors have the same problem one layer up, where the unit is an outcome: nine of 20 define it and all nine count a customer who stopped replying as a win. It is the same move as Clay's credit tables and the same move as the $100 and $200 tiers that all quote 5x and 20x: the multiplier does the number's work without doing the number's damage.