AI credits: what one costs and how to budget

What an AI credit is

An AI credit is the unit a vendor charges for AI work, and there's no standard. We've seen three meanings, and only the first is already in dollars.

Tokens vs credits

Tokens sit under every credit, and the model provider bills them. A token is roughly 4 characters or 0.75 words of English. For most models on OpenAI's standard price list, output tokens cost 4 to 8 times as much as input tokens. A task that writes a lot costs more than one that mostly reads. A credit tied to tokens moves with that, and one priced per action doesn't.

model cost = input tokens x input price
           + output tokens x output price
credits    = model cost / dollars per credit

Our example, with made-up prices of $2 per million input tokens and $10 per million output: a task that reads 20,000 tokens and writes 2,000 costs $0.06. At GitHub's $0.01 a credit, that's 6 credits. At Netlify's 180 credits per dollar of model use, it's about 11 (our arithmetic).

Why AI tools moved to usage-based pricing

Usage-based pricing means you pay for the AI work you use, and credits are how most tools count it. We think the reason is that one task can cost many times more than another, and a flat price can't cover both. Agents widen the gap. Anthropic found they use about 4 times more tokens than chat, and multi-agent systems about 15 times more.

GitHub moved Copilot plans to credits on June 1, 2026 because a quick chat question and a multi-hour autonomous session could cost a user the same, it said on April 27. Annual Pro and Pro+ plans keep the old pricing until they expire.

The trouble is that a credit hides the price until you do the math, and the bill grows with your usage. In Zylo's January 29, 2026 survey of 218 IT leaders, 78% reported unexpected charges tied to consumption-based or AI pricing in the prior 12 months.

Turn each credit into dollars

You need two numbers per vendor: what one credit costs you, and how many credits a task takes. If you expect to go over the allowance, use the extra-credit price.

dollars per credit = price paid / credits bought
cost per task = credits per task x dollars per credit

Salesforce has the cheapest credit in that table, and its actions still cost more. It charges 20 Flex Credits per action, about $0.10, while Microsoft charges 5 credits per agent action, about $0.04 (our arithmetic for both). So we'd compare what a whole task costs.

Estimate a month before you buy

1. Pick 3 to 5 real tasks from last week (our suggestion), not demo prompts. If you only try chat questions, agent work will look far cheaper than it is.
2. Run each in the trial and note the credits it used. If the vendor doesn't show that, ask, and count it as a red flag (see below).
3. Multiply each task's credits by how often you'd run it in a month, and add them up.
4. Add a buffer. We'd use 30%, because a trial rarely looks like a real month.
5. Convert to dollars and compare with the plan's allowance and overage rate. If you land over, price the next plan up.

Our example, with made-up numbers:

Task A: 40 credits x 20 runs   =   800
Task B: 120 credits x 8 runs   =   960
Task C: 15 credits x 100 runs  = 1,500
Total                          = 3,260
Plus a 30% buffer              = 4,238
At $0.01 a credit              = about $42

After the first month, compare real usage with your estimate. Then add the AI line to your weekly operations review and watch it like the cloud bill, as in AWS cost monitoring and tracking OpenAI and Anthropic API spend.

When the credits run out

For a trial, we'd take the hard stop. For anything your team relies on, we'd take automatic overage with alerts, since a stop mid-week can halt work. Overage without alerts is how surprise bills happen.

Cursor ran into this in 2025. It moved Pro to a $20 monthly pool billed at API rates, and some users were charged past it without realizing (TechCrunch). Cursor then offered refunds.

Whatever you pick, get alerts before the allowance runs out. HubSpot notifies at 75%, 85% and 90%. If you need a hard cap, bring your own provider key and set a spend limit with the provider. Anthropic lets you set a monthly one on its Billing page.

Red flags in credit pricing

None of these is a dealbreaker alone. We'd still get each answer in writing before paying.

• Ask when credits expire. A monthly reset is normal, but prepaid credits vary: Anthropic's purchased API credits expire one year after purchase and are non-refundable, while Netlify's auto recharge and pack credits have no set expiration date.
• If the vendor can't tell you what an extra credit costs, you can't budget past the allowance. GitHub, Microsoft and Netlify all publish theirs.
• You need to see which feature used the credits, or you can't tell what's eating the budget. HubSpot's Usage & Limits tab breaks credits down by feature, and Cursor added usage-limit visibility to its dashboard as part of its 2025 apology.
• Check what's switched on the day you sign up, like GitHub's default extra usage above.

One more we'd ask: real credits per task from a customer doing work like yours. If they can't show you that, your estimate is a guess.

How Since does credits

Since is a team of AI agents led by Grant. They read the sources you connect and report back, a fix comes as a draft pull request, and nothing merges or deploys unless you do it.

Credits are what that AI work uses, and each check-in or question you ask takes some. AI is included on every plan, so there's no model key to set up. Each plan comes with a monthly allowance, and inside it you pay nothing extra. Past it on Pro or Studio, extra credits are billed at the price per 1,000 in the table below (pricing).

• There's no spend cap on Pro or Studio. Work keeps going past the allowance, and extra credits go on a separate monthly invoice. Work stops at the allowance only when a payment is overdue.
• If you want a hard stop, owners and admins on paid plans can switch on "Use my own AI provider" and connect their own provider account. The provider bills that usage, so it uses no credits, and its spend limit is your cap (BYOK vs hosted AI). Free can't bring a key.
• The allowance resets on the 1st (UTC) and doesn't roll over. There are no top-up packs.
• There's an in-app banner at 80% and 100% of the allowance. The owner gets an email when it runs out, and again each time that month's extra credits pass another multiple of the plan price.
• Failed calls use no credits. Neither do World State watches, which follow public sources and have their own limits.
• On Studio and Enterprise, Liz, who covers costs, also reads your own OpenAI and Anthropic spend. That spend isn't counted in Since credits (tracking that spend).

Questions

What are AI credits?

AI credits are a unit each vendor defines to meter AI use. One might stand for dollars of model cost, a rate per 1,000 tokens or part of an action's price. There's no standard, so convert it to dollars before you compare.

What is the difference between tokens and credits?

Tokens are what the model provider bills. Credits are the vendor's own unit on top, tied to token cost at some vendors, as at GitHub, and to actions at others.

How do I convert tokens to credits?

Price the tokens first: input tokens times the input price, plus output tokens times the output price. Then divide by what one credit is worth. With made-up prices (our example), 6 cents of tokens is 6 credits at GitHub's 1 cent a credit.

How much does one AI credit cost?

Among the vendors we checked, one credit costs from half a cent for Salesforce Flex Credits to 1 cent for GitHub Copilot. At Since, extra credits are 2 dollars per 1,000 on Pro and 1.50 dollars on Studio (pricing). Each plan's monthly allowance is included in its price.

Do AI credits expire or roll over?

Usually the monthly allowance resets and unused credits are gone, as GitHub's are on the 1st. Prepaid credits can last longer: Anthropic's API credits expire after a year. Ask for the expiry in writing.

What happens when you run out of AI credits?

Usually one of three things: work stops, extra usage bills automatically, or you switch on pay as you go. At Since, Free stops until the 1st (UTC) or an upgrade, and Pro and Studio bill extra credits automatically (pricing).

What is usage-based pricing for AI tools?

You pay for the AI work you actually use, usually counted in credits. Vendors moved to it because one task can cost many times more than another, as GitHub said when it moved Copilot to credits.

Related guides

• Bring your own key (BYOK) vs hosted AI: which to chooseBring your own key (BYOK) means an AI tool runs on your own OpenAI or Anthropic LLM API key. When it beats hosted AI, and how to set it up safely.
• How to track OpenAI and Anthropic API usage and costsTrack OpenAI and Anthropic API usage and costs with what the providers already give you: usage pages, cost APIs, spend limits and a weekly LLM cost review.
• AWS cost monitoring for startups: a 15-minute weekly reviewTurn on free AWS cost alerts, read Cost Explorer by service each week and check the common leaks: NAT gateways, idle volumes, log retention and idle RDS.