Ad-supported AI coding agents in 2026: who pays, and what they can see

A free AI coding agent is either an open-source client you point at your own API key, or it runs on somebody else's money. Somebody else's money always has a shape, and that shape decides what the tool may do with your project.

The usual label is ad-supported, and it covers three different arrangements that expose you to different things. Two of them are not advertising at all.

Three products, one label

1. In-session ads

The advertisement is rendered inside the coding tool, usually in the waiting line while the model thinks. The advertiser is buying a developer's attention during work. Because ad selection happens in the same process that holds your repository, the vendor has to decide how much of your session goes out to target and measure it. That decision is the product.

2. A free daily grant

The vendor gives every account a fixed allowance per day and eats the inference bill as a customer acquisition cost. No advertiser is involved. This is the most comfortable model to use and the least stable to depend on, because an allowance funded by a marketing budget is a line item, and line items get cut.

3. A rewarded offerwall

You choose to complete an advertiser action outside the tool, usually a signup form or a survey in a browser tab, and the payout for that completed action becomes usage credits. This is the model mobile games have run on for fifteen years. The advertiser is paying for the completed action itself, so it has no reason to want your session and no route to it.

What happened to ad-supported coding in 2026

Most comparison articles still describe the state of things in early 2026. This is what changed.

Read as a sequence, those five rows say something specific. The in-session ad model was tried by the best-resourced company in the category and abandoned in under six months. What replaced it, the vendor-funded daily grant, then started being rationed eight weeks later. And the free tier at the largest vendor of all lost its strong models first, which is exactly what you would predict from the cost of a turn.

Can it see my code?

The answer differs per funding model rather than per vendor. The table shows what each arrangement structurally requires, before any company's policy is applied on top.

Two details from 2026 show this. AdGuard's analysis of Kickbacks.ai found that its private mode sends a pseudonymous identifier, a truncated IP, OS and editor version and an approximate location, while its boosted mode additionally processes recent prompts and AI responses on the vendor's servers along with open-file extensions and repository hints, with credential stripping described as best effort rather than guaranteed. That is what targeting attention inside a working session takes.

The second: on some tools the free mode is where the training terms live. A tool can truthfully say no code is shared with ad partners while its free tier still opts your threads into model training. Marketing pages tend to answer only the first question.

What one coding turn actually costs

Ad funding is only as plausible as the bill it has to cover. Below is the cost of a single agent turn, priced through the engine Clixad meters with, so these are the numbers we actually charge.

A "turn" here is defined as 20,000 input tokens and 1,500 output tokens: one round trip where the agent has a few files and some conversation in context and replies with an edit. Credits are Clixad's unit, where 100,000 credits is $1 of list value.

The same turn costs about forty times more on the frontier model than on the cheap one. A display ad in an editor earns single-digit cents per thousand impressions. That funds a lot of Gemini 2.5 Flash-Lite and almost no Claude Opus 5.5, which is why free tiers under pressure shed their strong models first, and why Sourcegraph's stated reason for dropping ads was a wave of frontier model launches.

One turn is right for comparing models and wrong for planning work: an agent makes about a dozen model calls per instruction and resends the whole conversation on each. The follow-up post prices a whole task instead.

What a rewarded action actually pays

The offerwall model works on different arithmetic: a completed signup or survey is worth a hundred to a thousand times a display impression. It also has two uncomfortable properties.

The payout varies enormously from one offer to the next. From our own ledger: completed offers on our CPX wall have credited from 9,000 to 114,000 credits on completion, with a median of 51,000: 26 completions recorded in our ledger between 27 July 2026 and 12 September 2026, as of 14 September 2026.

The largest paid more than twelve times the smallest. The ledger does not record how long an offer took, and we quote no figure for our other offerwalls.

Being screened out is the normal case, not an error. The most common outcome of starting a survey is that it ends part way through with nothing credited, or at most a small bonus, because you did not match the panel the advertiser was buying. Every amount above is what lands on completion.

For scale against the table in the previous section: one completion at the median is about 130 turns on Gemini 2.5 Flash-Lite, about six on Claude Sonnet 5.5, or about three on Claude Opus 5.5.

Five questions before you trust one

Each one is answerable from a vendor's public pages.

1. Where is the ad rendered? Inside the tool, or in a separate page you open on purpose. This single fact determines your exposure.
2. What leaves the machine to target it? Look for the specific list, not the reassurance. "We do not sell your data" is not a list.
3. Does the free mode also grant training rights? Check the free tier's terms, not the general privacy policy. This is the one that most often differs, and a later post follows what happened to it in 2026, when the largest free tier in the category was opted in by default.
4. Is the free allowance revenue or marketing? If no advertiser or customer is paying for it, it is a budget line, and 2026 has now shown twice what happens to those.
5. Can you see the price of a turn before you spend it? A tool that meters in a unit you can inspect is one you can plan around. A tool with an opaque "requests" quota is not.

Where Clixad sits, and what it does not solve

Clixad is a terminal coding agent that reads your project, edits files and runs commands. It is funded by rewarded offerwall completions, the third model, and also by a sponsored line in the terminal and one-time credit packs. For the offerwall you open a page in your browser when you want credits, complete a form or a survey, and the payout lands in a wallet the agent spends against. Advertisers never see your code, your prompts or the name of your project: the offerwall is a page in your browser, and the payout the advertiser buys is a completed form. New accounts start with 25,000 credits, and buying credits is never required.

The limits: you are paying with your time and your survey answers, which is a real price even though it is not a subscription. The offerwall provider collects its own data about you, on its own terms, and that is a second privacy question we cannot answer on its behalf. Screenouts will waste some of your minutes. The sponsored line, added in August 2026, sits above the input box while a model call runs, and two more places on screen can show an ad too. Each is marked as an ad and fetched from an ad network by your machine, which means that network sees your IP address; each is chosen on a fixed category and never on anything you typed, and CLIXAD_SPONSOR=0 stops those requests. That is a second ad-funding model in one product, and the framing above applies to it too. And nothing here is appropriate for code you are contractually obliged to keep confidential.

What the model does buy you is a free tier that is not a trial. The margin on a completed offer is the business, so a user who never pays a cent is profitable rather than a cost tolerated until the next funding round.

FAQ