Every free AI coding tier is advertised in a unit that cannot be converted into work. Requests, premium requests, completions and daily grants all count interactions or dollars the vendor will not price for you. Meanwhile one ordinary agent task is about a dozen model calls, and because each call resends the whole conversation, it sends 348,600 input tokens
rather than the 144,000 a naive twelve-times-the-first-call estimate gives you. So the honest comparison needs a unit that survives an agent loop. Tokens at the model's published price, or credits at a published rate. Anything else is a number somebody chose.
Every comparison of free AI coding tools ends up as a table of limits: so many requests a month, so many premium requests, so many per minute, a daily grant that resets at midnight. The tables are accurate. They are also unusable, because not one of those numbers can be turned into the thing you want to know, which is how many actual jobs you get to finish before the tool stops.
Vendors did not invent that gap to be evasive. Measuring this got genuinely harder the moment coding assistants became agents, and the units on the pricing pages never caught up.
Five units, none of them work #
- Request
- One thing you asked for. GitHub's documentation defines it as "any interaction where you ask Copilot to do something for you", and states that for agentic features only the prompts you send are counted, while the tool calls the agent makes on its own are not. The quota therefore measures your typing, not the model's work.
- A request weighted by a per-model multiplier. The multiplier is the interesting part: it exists because a request has no fixed size, so the vendor has to correct for the model you picked. It corrects for the model but not for the length of the job. GitHub documents one Copilot code review as consuming 13 premium requests, which is a fair price and also an admission that the unit does not hold.
- Daily grant
- An allowance of model usage that refreshes each day, usually stated in a currency the vendor does not publish a conversion rate for. It is the only unit here that scales with real cost rather than with interactions, and also the one most easily reduced, because it is set by whoever is funding it. Amp cut its free daily allowance for more users in May 2026.
- Token
- The unit providers actually bill in. It is exact and it is the only one that survives contact with an agent loop. A raw token count still hides the price difference between models, though, where the same token can cost four hundred times more depending on which one you sent it to.
- Credit
- Tokens already priced at the model's rate, so one number covers both how much work you did and which model did it. Clixad meters in credits, at 100,000 credits to $1 of list value. This is the unit we chose and we are not neutral about it, so treat the rest of this article as the argument for it rather than as a survey. A credit is only as useful as the published conversion rate behind it.
The pattern across the first three: they count what you did. Since 2025 the agent does most of the work between your instructions, so a unit anchored to your keystrokes drifts further from the bill every time agents get more autonomous.
One task, twelve calls #
Take an instruction an agent gets a hundred times a day: find why this test fails and fix it. The agent reads the failing test, greps for the function, reads two files, makes an edit, runs the suite, reads the new error, edits again, runs it again, and reports back. On a request-counted plan that is one request. In the provider's logs it is roughly a dozen calls.
The second thing that happens is the one nobody prices in. A model has no memory between calls, so every call carries the entire conversation up to that point. Each file the agent reads and each command it runs is appended, and then it is sent again on the next call, and the one after. The context does not accumulate in the model, it accumulates in your bill.
Here is the task written out. It starts at 12,000 input tokens, which is a system prompt, a project file and your instruction. Every step adds 3,100 tokens of tool output and previous reply, and every step generates 600 tokens back. Twelve steps.
| Step | Input tokens sent | Sent so far |
|---|---|---|
| 1. plan the fix | 12,000 | 12,000 |
| 4. read the second file | 21,300 | 66,600 |
| 8. run the suite again | 33,700 | 182,800 |
| 12. report back | 46,100 | 348,600 |
The twelfth call sends 46,100 tokens to answer a question you asked once. Cumulatively the task costs 348,600 input tokens and 7,200 output tokens. If you had estimated it as twelve calls the size of the first one, you would have said 144,000, and you would have been under by a factor of about two and a half.
This is why "how many requests do I get" has no useful answer. The same request is a tenth of the bill on a one-shot question and forty times that on a debugging session, and the person who wrote the quota knows this, which is why the quota is not stated in tokens.
Caching changes the size of the effect but not its direction. Providers charge less for a repeated prefix, so a well-cached agent pays a discount on the resent context rather than avoiding it. The figures here are list price with no cache assumed, which is the number your wallet has to be able to cover in the worst case.
What that task costs on fourteen models #
Below is the same twelve-step task priced through the engine Clixad meters with, on every model in our catalogue rather than a flattering subset. Each call is metered and rounded on its own, the way the gateway actually bills it. The second column is the estimate you would have reached by pricing twelve calls at the size of the first.
| Model | Credits for the task | If context never grew |
|---|---|---|
| GPT-5 Nano | 3,051 | 1,512 |
| Gemini 2.5 Flash-Lite | 5,664 | 2,592 |
| DeepSeek V4 Flash | 7,629 | 3,336 |
| Gemini 3.1 Flash-Lite | 14,697 | 7,020 |
| Kimi K2 | 32,297 | 14,796 |
| Kimi K2.7 Code | 43,002 | 19,980 |
| Claude Haiku 4.5 | 57,690 | 27,000 |
| Gemini 3.6 Flash | 86,538 | 40,500 |
| GPT-5.2 | 106,632 | 52,920 |
| Claude Sonnet 5 | 115,380 | 54,000 |
| Kimi K3 | 173,070 | 81,000 |
| Claude Opus 5 | 288,450 | 135,000 |
| GPT-5.5 | 293,850 | 140,400 |
| GPT-5.2 Pro | 1,279,531 | 635,040 |
Two things fall out of that table. The undercount runs between roughly two and two and a third times across every model, so it is not a quirk of one price list. And the spread from the top row to the bottom one is more than four hundred times, on a task defined identically for both. That is why free tiers are quoted in requests. A request can be sold as one thing. This cannot.
The cheapest model depends on the shape of your work #
Our model catalogue is ordered by what a single conversational turn costs, defined in the previous post as 20,000 input tokens and 1,500 output tokens. By that measure Claude Sonnet 5 costs 8,250 credits and GPT-5.2 costs 8,401, so Sonnet is the cheaper of the two.
On the twelve-step task the order reverses: GPT-5.2 comes to 106,632 credits and Claude Sonnet 5 to 115,380. Nothing changed about the models. An agent task is input-heavy and output-light, GPT-5.2 charges less for input and more for output, and at this ratio that trade works out in its favour. Sort the same catalogue by a chat-shaped workload and it flips back.
I find this the most useful thing in the article, because it kills the idea that there is a ranking to memorise. There is a ranking per workload shape. Any tool that shows you one league table of models by price is showing you the shape its author had in mind.
What one offerwall completion covers #
Clixad funds usage with rewarded offerwall completions, so the same conversion applies to us and it is only fair to run it. These are payouts observed on our wall in Germany, credited on completion:
- a 3 minute action: roughly 40,000–42,000 credits
- a 6 minute action: roughly 45,000–107,000 credits
- a 16 minute action: roughly 300,000 credits
Duration barely predicts the payout, which is why those are ranges. The advertiser is pricing the respondent rather than the minutes, and two six-minute surveys can differ by more than 2x. Take the low end of the shortest one, 40,000 credits, against the table above: it covers about seven of these tasks on Gemini 2.5 Flash-Lite, roughly one on Kimi K2, and about a seventh of one on Claude Opus 5.
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 and pays nothing, because you did not match the panel being bought. That is how offerwalls work everywhere and a product that presents it as a failure is describing the wall inaccurately.
We also cap what one free account can earn per day, at 500,000 credits. Publishing your own ceiling is the least fun paragraph to write in an article criticising other people's limits, but an article arguing for convertible units would be worthless if ours were not.
A completion is not the only shape ad funding takes, either. The other live one puts the advertisement inside the session rather than in a browser tab, which changes what the funding needs to see — a later post compares the two directly.
Converting any quota into work #
You can do this arithmetic on any tool, including ours, from public pages and one afternoon of use. The order matters: each step is cheap only if the previous one is done. A later post runs it across the six free tiers that include model access, and finds that four of them grant a unit the arithmetic cannot be applied to at all.
Find out which model is actually serving you on the free tier. Not which models the vendor sells. Free tiers get restricted to the cheap end first, which Google did to the Gemini CLI free tier in March 2026, and whichis one of the defaultsa vendor can move without touching a single published limit. - Price one task, not one message. Give the agent a job you would really give it and count the model calls it makes. A dozen for something ordinary is a reasonable planning number. - Multiply the context growth back in. If you estimate from the first call, double it. That is the correction the table above measures, and it is the step almost everyone skips. - Convert the quota into that unit. If you cannot, because the quota is in requests and requests have no size, you have learned the thing worth knowing about the offer. - Ask what happens at zero. A tier that degrades to a smaller model is a different product from one that stops, and both are different from one that silently truncates your context to fit.A later posttakes that question apart, including what an abandoned task has already cost by the time it stops.
What Clixad does, and where it also runs out #
Clixad is a terminal coding agent funded by rewarded offerwall completions rather than by a subscription. It meters in credits at a published rate, it shows you the price of a turn before you spend it, and the wallet is checked before generation starts rather than after, so a request you cannot afford is refused instead of being half-served and clawed back. New accounts start with 5,000 credits.
Which, by the numbers above, does not cover one full twelve-step task on Gemini 2.5 Flash-Lite. It is a taste, not a tier, and calling it anything else would be the exact move this whole article objects to. The daily earning cap is real too. On Claude Opus 5 the most a free account can draw in a day is under two tasks of the kind priced here.
What the model does fix is the unit. Whether that is worth completing a survey for is a reasonable thing to decide against, and you can decide it from published numbers rather than from a quota you would have to reverse-engineer.
FAQ #
How many model calls does one AI coding task use?
Far more than one, and the number is set by the agent rather than by you. A single instruction like "find why this test fails and fix it" sends one call to plan, one per file read, one per edit, one per command run and one to summarise, which lands somewhere around a dozen for an ordinary task and considerably more for a hard one. Every one of those calls is billed, and every one of them carries the whole conversation so far.
Why do free AI coding tiers count requests instead of tokens?
Because requests are the unit the user can see. Tokens are the unit the provider bills in, but a quota stated in tokens invites the question of what a token costs, and that question has a different answer for every model. Counting requests keeps the number small and memorable at the price of it no longer measuring anything. The tell is the multiplier: as soon as a vendor has to weight a request by which model served it, the request has stopped being a unit.
Do tool calls count against my quota?
On a request-counted plan, usually not, and that is the part people misread as generosity. GitHub's documentation says that for agentic features only the prompts you send count, while the actions the agent takes on its own do not. The model calls behind those actions still happen and still cost the vendor money. The vendor is absorbing them, which means it retains every reason to cap them somewhere you cannot see.
Is a free AI coding tier enough for real work?
It depends almost entirely on which model the free tier lets you point at, not on how large the quota sounds. The same task priced through our engine costs 3,051 credits on GPT-5 Nano and 288,450 on Claude Opus 5, a spread of roughly ninety-five times. A free tier that is generous in requests but restricted to a small model is a genuinely usable thing for scripting and small fixes. A free tier that offers a frontier model in small numbers runs out during your second task.
How do I compare free AI coding tiers fairly?
Convert every offer into the same unit before comparing, and use one you can verify: tokens at the model's published price, or credits at a published rate. Then ask what one finished task costs in that unit, not what one interaction costs. A quota you cannot convert is not a comparison, it is a number the vendor chose.
Sources #
- GitHub,
Requests in GitHub Copilot, for the definition of a request, the treatment of agentic tool calls, the per-model multiplier and the cost of a code review. Read 1 August 2026. - Sourcegraph,
Amp Free Is Ad-Free, for the 23 May 2026 note on pausing or reducing the free daily allowance for more users. - Google,
Service update: mitigating abuse and prioritizing traffic, Gemini CLI discussion #22970, effective 25 March 2026, for the restriction of the free tier to Flash models. - Per-task credit prices: Clixad's own pricing engine,
packages/pricing
, at the model catalogue read from OpenRouter on 26 July 2026. The task shape is our own worked example, stated in full above so it can be recomputed. - Offerwall payouts: amounts observed on Clixad's own wall in Germany, credited on completion, July 2026.