The Hidden Costs of AI Tools (Credits, Overages, Fair-Use Caps)

TL;DR: The biggest AI tool hidden costs aren't secret fees. They're mechanics on a different page from the price: included credits that expire, allowances with a variable portion, overage billed after the fact, fair-use conditions on "unlimited" plans, and prices shown before tax. We read the vendors' own pages on September 16, 2026 and quote them below, then list our own caps.

Almost nobody reads the pricing page past the big number. That's where AI tool hidden costs live. Not in fine print designed to trick you, but in the paragraphs that explain what "included", "unlimited" and "credits" actually mean.

This post reads those paragraphs for you. Every vendor quote below comes from the vendor's own page, fetched on September 16, 2026. Where a vendor's pages disagree with each other, we say so. At the end, we hold our own pricing to the same standard, because a post about other people's caps that hides ours would be exactly the problem it describes.

Why does an AI tool cost more than its sticker price?

Because the sticker price buys an allowance, not an outcome. Most AI subscriptions now include a set amount of model usage, and how fast you burn through it depends on which model you pick, how long your conversations run, and which features you use.

Cursor states the model plainly on its pricing page: "Every plan includes a set amount of model usage." Its docs then explain the variable part: "Since different models have different API costs, your model selection affects how quickly your included usage is consumed."

Anthropic describes the same dynamic for Claude from the other direction. Its help center says usage "is affected by several factors, including the length and complexity of your conversations, the features you use, which Claude model you're chatting with, and the effort level you've selected." It also notes that "your usage of all different Claude product surfaces (claude.ai, Claude Code, Claude Desktop) counts towards the same usage limit."

So two people on the same plan can have very different months. One uses a light model for quick questions. The other runs long agent sessions on a frontier model. Same price, different amount of work before the limit arrives. If you want to compare what a unit of usage is worth across tools, what an AI credit is actually worth does that math.

Do AI credits expire if you don't use them?

Often, yes. This is the most common cost people never notice, because it shows up as value lost rather than a charge.

GitHub is explicit about Copilot: "Included AI credits do not carry over between months. Unused credits are forfeited, and your allowance resets to the full monthly amount at 00:00:00 UTC on the first day of each calendar month." It adds that the reset date "is fixed and does not change based on your subscription billing date", so if you subscribe late in a month, your first allowance resets within days.

Lovable has several expiry rules for different credit types. Its pricing FAQ says: "If unused, monthly plan credits expire two (2) months after they're issued, and annual plan credits expire one (1) month after your annual period ends. Top-up credits last twelve (12) months from purchase." Daily grants are shorter still: "Included grants of daily build credits expire at the end of each day". And on refunds: "Credits aren't refundable or redeemable for cash or any other value."

None of this is unreasonable. It is, however, a cost. If you buy a top-up in a busy month and your workload drops, the unused balance has a clock on it.

Source: vendor pages linked below, read September 16, 2026.

What is a "flex" or variable allowance?

It's an allowance whose size the vendor can change. GitHub introduced one for Copilot, and it's worth understanding because other vendors may copy it.

Copilot's individual plans now include two layers. GitHub describes base credits as "included with your plan subscription each month" and says "they never change." On top of that sits a flex allotment, which GitHub describes as "a variable part of your included usage; it is designed to adapt as the economics of AI evolve, including model pricing, new models, and improvements in efficiency."

On September 16, 2026, GitHub's table showed:

Look at the Max row. Half of the included usage is the part GitHub describes as variable. That doesn't mean it will shrink. It does mean the number you compare today isn't guaranteed to be the number you get later. When you budget, count the fixed part as what you're buying.

Copilot also shows how a single feature can cost more than it appears to. For subscribers still on GitHub's older request-based billing, its legacy docs say: "Starting June 1, 2026, Copilot code review will have a model multiplier of 13." Each review drew down thirteen requests, not one. That page applies only to some annual subscribers, but the lesson carries: check whether any feature you rely on has a multiplier.

How do AI overage charges work?

Usually as metered billing that starts when your included usage ends. The bill arrives afterwards, which is what makes it feel hidden.

Cursor's pricing FAQ says: "On-demand usage allows you to continue using models after your included amount is consumed, billed in arrears." Its docs add that on-demand usage continues "at the same API rates with pay-as-you-go billing", and they give unusually candid guidance on what real use costs. Under "How much usage do I need?", the docs list "Daily Agent users : Typically $60–$100/mo total usage" and "Power users (multiple agents/automation) : Often $200+/mo total usage". Cursor Pro is listed at $20 a month on the same site. If you're a daily agent user, the page itself tells you to expect a bill several times the sticker price.

Teams on Cursor have one more line item. The docs say that on Teams and Enterprise plans, "third-party model requests include a Cursor Token Rate of $0.25 per million tokens", which "applies on top of model API pricing for included usage, on-demand usage, and BYOK usage." Cursor's own models are exempt.

GitHub's overage is budget-based. When Copilot credits run out, you can set a budget for more: "GitHub AI Credits draw down your budget at a fixed rate: 1 AI credits = $0.01 USD, so a $10 budget covers 1,000 AI credits." GitHub also warns that "additional usage may be capped", in which case you'll need to pay what you've already used before continuing.

Anthropic gives Claude subscribers three options when they hit a limit: "you'll need to wait for it to reset, upgrade your plan, or purchase usage credits."

Is "unlimited" AI really unlimited?

Rarely without conditions. The word usually comes with a footnote, and the footnote is the policy.

ChatGPT's pricing page, read on September 16, 2026, attaches the same line to several plans: "Unlimited text chats subject to abuse guardrails." The feature table marks many entries "Unlimited*", and the asterisk resolves to: "*Usage must be reasonable and comply with our policies". The page also notes that the Go plan "may include ads", which is a different kind of cost, paid in attention rather than money.

Perplexity shows how "unlimited" and "limits" can sit on the same page. Its help article describes Education Pro as including "unlimited Pro Searches". The plan comparison table in the same article lists Pro Searches for Education Pro as "Weekly limits", with the note "(average use)". Both statements were on the page when we read it. We can't tell you which one Perplexity enforces; we can tell you to ask before you rely on either.

Anthropic avoids the word for Claude subscriptions. Its help center has a section titled "How can I get unlimited usage?" that answers by pointing to purchasable usage credits, and says only accounts on a usage-based Enterprise plan have usage "based on consumption."

A fair-use clause is not a trick. Every flat-rate AI product has one, because a single automated script can cost more than a thousand normal users. The hidden cost is uncertainty: you don't know where "reasonable" ends until you reach it. For a deeper look at which tools publish a real uncapped tier, see prompt tools with unlimited prompts and no monthly caps.

What other costs don't appear in the headline price?

Three come up again and again.

Tax. Cursor says: "All prices are exclusive of any applicable taxes." Anthropic's nonprofit pricing page notes: "Prices shown don’t include applicable tax." Depending on where you are, that's a real percentage on top.

Seat minimums. ChatGPT's pricing FAQ says "Business plans are available starting at 2 users." Anthropic's discounted nonprofit Team plan lists a "Minimum 2 seats". A solo buyer on a team plan pays for a second seat.

Invoicing and payment methods. Cursor points organisations "that need invoicing, pooled usage, or advanced security" to Enterprise. OpenAI says "For ChatGPT Enterprise, please contact sales" for options "such as invoicing." If your finance team needs an invoice rather than a card charge, the plan you can buy may be pricier than the plan you want. (If you're trying to get an employer to pay, getting AI tools reimbursed covers the paperwork side.)

How do you audit an AI subscription for hidden costs?

Read five things before you pay, and check one dashboard after. This checklist works for any vendor:

AI SUBSCRIPTION COST CHECK
Tool: ____________   Plan: ____________   Date checked: ________

1. INCLUDED USAGE   What exactly is included? (messages, credits, $ of usage)
                    Is any part labelled variable, flexible or subject to change?
2. EXPIRY           Do unused credits roll over? When do they reset or expire?
                    Is the reset tied to my billing date or the calendar?
3. OVERAGE          What happens at the limit: block, slow down, or bill?
                    Can I set a hard spending cap? Where?
4. "UNLIMITED"      What does the footnote say? Link to the usage policy.
5. EXTRAS           Tax included? Seat minimum? Invoice available on this plan?

AFTER 7 DAYS        Usage so far: ______   Projected month: ______
                    Still the right plan?  yes / no

Then do the simple things that prevent most surprise bills. Set a spending limit wherever one exists. Turn on usage alerts. Check the usage page weekly for the first month. And if your usage is steady, prefer a flat plan; metered pricing only wins when your usage is genuinely spiky. Teams buying for many people should read prompt tools for customer support teams, where per-seat costs multiply fast.

What are Prompt Architects' own limits?

Here they are, in the same plain terms. Figures are from our pricing page and FAQ, at the time of writing (September 16, 2026, during a launch promotion).

A few things we want to be upfront about:

• Advanced "unlimited" has conditions too. Our terms of service prohibit using "automated systems or bots to access the Service beyond normal usage" and prohibit attempts to "Circumvent usage limits or restrictions". That's our fair-use clause, and we'd rather you read it here than find it later.
• Pro is a monthly count. If you enhance more than 200 prompts in a month on Pro, you need Advanced.
• Team shares one quota. Teammates draw from a shared pool, and the owner manages seats and billing.
• Free limits can change. Our terms say of the Free plan: "Features and limits may change at our discretion."
• Paid prices can change, with notice. Our terms say: "We reserve the right to change pricing with 30 days' notice to active subscribers."
• No API key and no metered model bill. Every plan includes built-in AI for enhancement, so there's no separate model invoice behind the subscription.
• Refunds. A 14-day money-back guarantee on first-time paid purchases; outside that window, no refunds for partial billing periods unless the law requires one.

We don't sell credit top-ups, so nothing expires unused, and there's no overage billing. The trade-off is that Pro has a hard monthly count rather than a soft one. If your usage is unpredictable, that's worth knowing before you choose.

So which pricing model hides the fewest costs?

The one whose limits you can state in a sentence. A flat plan with a published count is the most predictable. A metered plan with a hard spending cap is safe. A plan with a variable allowance, expiring credits and a vague reasonable-use footnote isn't necessarily a bad deal, but it's the one where your real cost is hardest to know in advance.

Before you buy anything, run the checklist above. Before you renew, compare the month's actual usage with what the plan included. Most AI tool hidden costs disappear the moment you read the second paragraph of the pricing page.

Sources, all read September 16, 2026: GitHub Docs, usage-based billing for individuals · GitHub Docs, requests in Copilot (legacy) · Cursor pricing · Cursor Docs, models and pricing · ChatGPT pricing · Claude Help Center, usage and length limits · Claude for Nonprofits · Lovable pricing · Perplexity Help Center, plan comparison · Prompt Architects pricing · Prompt Architects terms