Free tiers don't have a price, they have defaults

Every comparison of free AI coding tools is a table of quotas, and a previous post here argued that the quotas cannot be converted into work. This one is about the rest of the offer: the terms that are not printed anywhere because they were never presented as terms. They were presented as settings, with a value already filled in.

That distinction decides more than it should. A price has to be agreed. A default only has to be announced.

A default is a price nobody quoted you

When a paid plan changes, you get an email, a renewal date and the option to leave. When a default changes, you get a changelog entry. The mechanism is not sinister, it is just cheap, and cheap mechanisms get used more often.

It is also where the money is. A free tier's real cost to a vendor is not the quota, it is the quota multiplied by the model serving it, and only one of those two numbers is on the pricing page. Moving the other one is invisible.

The three axes, and the fourth

Which model serves you

The largest factor in what a free tier is worth, and the easiest to change without changing a published number. The same twelve-step agent task costs more than four hundred times as much on the most expensive model in our catalogue as on the cheapest. Moving a free tier down the catalogue cuts the vendor's bill by more than any quota reduction would, while every advertised limit stays exactly as printed.

How much you get

The quota, and the only one of the three that pricing pages state. It is also the one most visibly funded by somebody's marketing budget rather than by revenue, which is why it moves when that budget is reviewed rather than when costs change.

What happens to what you type

Whether your prompts, the model's outputs and the code around your cursor are retained and used for training. This is the axis that moved most decisively in 2026, and the one where the default differs by who is paying.

What happens at zero

The fourth, and the one you meet first. Running out can mean a hard stop, a silent move to a weaker model, metered overage, or an interruption that loses the task you were running. The previous post takes that one apart, including what an abandoned task has already cost by the time it stops.

What actually moved in 2026

Four dated changes, each from the vendor's own publication. Nothing here is a leak or an allegation. They were all announced properly, which is the point: this is what normal looks like.

The April change is worth quoting rather than summarising. GitHub's changelog says that from April 24 onward we will begin using interaction data, specifically inputs, outputs, code snippets, and associated context, from Copilot Free, Pro, and Pro+ users to train and improve our AI models. The opt-out is a dropdown in your Copilot settings, it is free, it takes about fifteen seconds, and GitHub states that once you opt out it stops collecting from that point forward. Two further details belong in the same breath, because leaving them out would make this less accurate than the source: GitHub says it does not use private repository content at rest for training, and a user who had previously opted out of data collection for product improvements stayed opted out.

So this is not a scandal. It is a well-run company changing a default, in public, with a working switch. What makes it worth an article is the shape of the change rather than its ethics. The plans that pay per seat were exempt. The plans that do not were opted in.

What a free allowance is worth

Here is the part that reframed this for me. Copilot Free is advertised at 2,000 completions per month plus limited chat and agent usage. The completions are the only part with a published number, so they are the only part that can be priced.

An inline completion is a small job: some surrounding code goes up, a line or two comes back. No vendor publishes the token shape of one, so the table below assumes 1,200 input tokens and 40 output tokens, states that assumption in its own markup, and prices 2,000 of them through the same engine Clixad meters with. The four models are a bracket, from the cheapest in our catalogue to a mid-sized one, rather than a guess at what GitHub runs.

Inline completions are served by small, fast models in practice, so the top row is the realistic one and the bottom row is an upper bound nobody is paying. A month of the advertised allowance is plausibly a quarter of a dollar of inference.

For comparison, on our own wall in Germany a 3 minute action has paid roughly 40,000–42,000 credits, credited on completion. One of those is worth more inference than that entire monthly allowance priced on GPT-5 Nano, and about the same priced on Gemini 2.5 Flash-Lite. Being screened out is the normal case, not an error: most attempts end part way through and pay nothing, which is how offerwalls work everywhere. Duration barely predicts the payout either, which is why that is a range and why our observed spread runs up to around 300,000 credits.

None of that says Copilot Free is a bad deal. It says the deal is not mainly about inference. Completions are cheap to serve, cheap enough that giving away two thousand a month is a rounding error against the value of being the tool a few million developers open by default, and against whatever the interaction data is worth to the people training the next model. The quota is the part you can see. It is not the part being traded.

Read the funder, not the policy

A privacy policy is a promise, and promises are revised with a changelog entry. A funding model is a constraint, and constraints are much harder to move. So the question that survives contact with 2027 is not what a vendor currently permits itself, it is what the money would make it want.

• A vendor that sells models has a direct use for your prompts and your accepted completions. It may not use them, and it may say so convincingly, but the incentive is structural and it points one way.
• A vendor that sells seats to your employer has a strong reason to treat your employer's code carefully and a weaker one to treat your side project the same way. That is precisely the split the April change made visible.
• A vendor whose free tier is marketing is funding it from a budget that gets reviewed. What gets cut is the quota or the model, and the announcement will be about sustainability.
• A vendor funded by a rewarded action has no use for your code, because the money arrives from an advertiser for a completed form rather than from anything in your editor. That is our arrangement, and the first post here works through what it does and does not protect, including the things advertising exposes that a subscription does not.
• A vendor funded by ads inside the session is a different case, and the earlier wording here was too generous to it. If the advertisement is selected while you work, something has to select it: FreeBuff's launch post states that prompts and messages may be analyzed to personalize ads. "Ad-funded" is therefore not one privacy story but two, split by where the ad is rendered — the comparison post takes that apart, and the shortlist post places both among the free tiers that include model access.

Our defaults, written down

It would be cheap to write all of that and not state our own, so: Clixad is a terminal coding agent funded by rewarded offerwall completions. Its defaults, as they stand today.

What happens to what you type. The CLI holds no provider key, so every model call goes through our gateway, and your prompt and the file contents the agent read pass through it in transit on the way to the model provider. That is the price of metering centrally and I would rather write it down than imply otherwise. We do not train on any of it, which is less of a virtue than it sounds: we have no model to train and no way to make money from your code. The offerwall receives an account identifier and pays for a completed action. It receives no code, no prompt and no project name, because the ad is a page in your browser rather than a panel in your editor.

Which model serves you. Every model in the catalogue is funded the same way. There are no locked tiers, so a stronger model costs more credits per turn rather than requiring a plan. The prices come from the provider's published rates and are pinned by tests to the figures on this site.

How much you get. New accounts start with 5,000 credits, which does not cover one agent task on most of the catalogue. A free account can earn at most 500,000 credits a day across every offerwall we run. Both of those are levers we control, and both are exactly the kind of default this article is about, which is why they are published rather than discovered.

If offerwall payouts fell far enough, the levers that would move are the daily cap and the margin, not the data question, because there is nothing on the data side to sell. That is not a promise about our character. It is the shape of the business, which is the only kind of assurance worth reading.

FAQ