How to get free AI coding credits in 2026

By Florian Rappold / / 6 min read
Short answer

Free AI coding credits come from four places: a provider trial balance, a student or startup programme, a vendor's free tier, or an ad-funded grant. The first two are one-time and carry an expiry date. The third is usually not denominated in credits at all. The fourth comes back every day.

The word credit now does four incompatible jobs. One vendor's credit is one message, another's is a fixed fraction of a dollar, and ours is metered against real token usage. Only the last kind tells you what it finishes.

The four sources

What decides whether a free balance is useful is whether the grant returns, more than its size.

Where free AI coding credits come from, and how each behaves over a month of real use.
Source Returns? Unit granted
Provider trial balance No, one time Dollars of raw API usage, with an expiry date
Student or startup programme No, one time A larger dollar balance, behind an eligibility check
Vendor free tier Yes, monthly or daily Messages, requests or sessions, rarely tokens
Ad-funded grant Yes, daily Credits metered against real token usage

The two "no" rows are where most published advice stops, and they cannot carry a project. A trial balance shows whether you like a tool, but there is no second grant, and the clock on the first is already running.

The expiry clock

The expiry terms decide whether a saved balance still exists when you need it. OpenAI's service credit terms make credits non-refundable and expire them one year after purchase or issuance, and that covers promotional credits they issue as well as credits you buy. Others are tighter: Framer expires top-up credits at the end of the calendar month they were bought in.

So do not save a trial balance for a serious project: a credit with a deadline is only worth what you spend before the deadline.

Does this expire is about the balance; does this return is about the arrangement. A grant that returns daily makes expiry almost irrelevant, which is why the recurring sources are the ones to plan around even when their numbers look smaller.

Four units, one word

In 2026 "credit" means at least four different things, and no conversion exists between them.

What one credit means, by vendor. There is no exchange rate between these rows.
Vendor One credit is Converts to work?
Base44 One message, 25 a month on the free plan No
GitHub $0.01 of usage; the free plan receives none Partly, once you know the token rate
Linear A dollar of a balance you top up, expiring after 12 months Partly, same caveat
Clixad 1/100,000 of a dollar of list value, metered on tokens Yes

Base44's row is the clearest case. Twenty-five message credits a month sounds like an allowance, but a request is not a unit of work: one ordinary agent task is about a dozen model calls, because the whole conversation is resent every round. Twenty-five messages is about two tasks.

The same applies to a tier counted in sessions, like FreeBuff's six one-hour sessions a day, and to Cursor's Hobby plan, which no longer publishes a figure at all. The routes to coding without paying all run into this.

What a recurring grant is worth

Our own grants, priced. The workload is the twelve-call agent task from the tier-limits post. One task costs 3,051 credits on GPT-5 Nano, 54,000 on Kimi K2.7 Code and 230,760 on Claude Opus 5.5.

Clixad's three grants converted into whole finished tasks. Computed by the metering engine and recomputed by a test. Earned grants are credited on completion.
Grant GPT-5 Nano Kimi K2.7 Code Claude Opus 5.5
Signup bonus, 25,000 credits 8 tasks 0 tasks 0 tasks
One completed CPX offer, 9,000 to 114,000 2 to 37 tasks 0 to 2 tasks 0 tasks
Daily maximum, 500,000 credits 163 tasks 9 tasks 2 tasks

The middle row is a range because payouts vary widely from one offer to the next, and the ledger does not record how long an offer took. What the ledger holds: 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.

Credits arrive on completion, and a screenout part way through a survey pays nothing or a small bonus while still taking the minutes. That is the most common outcome of starting one, on every offerwall. Weigh 500,000 a day against the attempts that pay little or nothing too.

The signup bonus is eight tasks on the cheapest model and none on the other two: the same grant is a week of work or nothing, depending on the row. And a full day of earning is two tasks on Claude Opus 5.5. Frontier models are not cheap here.

Correction, 19 October 2026. This paragraph said frontier models are not locked behind a tier. Since 4 October 2026 they are premium models, which run on credits from a completed offer or a purchase; the signup bonus and other bonus credits pay for budget models only.

Stacking them

These sources combine, so spend the perishable ones first.

Spend the trial balance now, on a cheap model. It expires, it does not return, and its purchasing power varies by more than 400x across a model catalogue. Spending it on a frontier model is how $5 becomes six tasks instead of several hundred.

Claim a programme balance if you are eligible, and treat it the same way. Student and startup credits are larger and behave identically otherwise.

Keep a free tier for the work it suits. An allowance counted in inline completions is genuinely generous for inline completion work, and useless for agent work.

Use a recurring grant for anything that has to still be running next month. This is the only category where the arithmetic above still holds in week four. Know what happens when the meter stops as well.

FAQ

Where can I get free AI coding credits?

From four places. Provider trial balances give a one-time amount, typically around $5, that expires. Student and startup programmes give a larger one-time amount if you qualify. Vendor free tiers give a recurring allowance, usually counted in messages, requests or sessions. Ad-funded agents give a recurring grant in credits, earned by completing an advertiser action. Only the last both renews and converts into finished work.

Do free AI credits expire?

Usually. OpenAI's service credit terms make credits non-refundable and expire them one year after purchase or issuance, promotional credits included. Framer's top-up credits expire at the end of the calendar month they were bought in. A trial credit saved for a serious project may never get spent.

Why can't I tell what my free credits are worth?

Because "credit" covers at least four incompatible units. A Base44 message credit is one message. GitHub prices an AI Credit at $0.01, and its free plan gets none. Linear sells a dollar balance that expires twelve months after purchase. Clixad meters credits against real token usage at 100,000 credits per US dollar of list value. Only a unit tied to tokens tells you what a balance finishes.

How many credits does one coding task actually need?

On our engine, one twelve-call agent task costs 3,051 credits on GPT-5 Nano, 54,000 on Kimi K2.7 Code and 230,760 on Claude Opus 5.5. A 25,000-credit signup bonus is eight tasks on the cheapest model and none on the other two. One completed CPX offer at the median, 51,000 credits, is 16 tasks on the cheapest model and none on most of the catalogue. The model you pick matters more than the tier you are on.

Are ad-funded credits worth the time they take to earn?

It depends on what your time is worth and which model you run. Credits arrive on completion, and being screened out of a survey part way through pays nothing or a small bonus while still costing the minutes; that is the most common outcome on any offerwall. Payouts vary widely: 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.

Sources

  1. OpenAI, Service credit terms, for credits being non-refundable and expiring one year after purchase or issuance. Summarised rather than quoted: the page refuses automated retrieval, so this is stated from OpenAI's published terms as reported by their help centre rather than from a verbatim read. Checked 13 August 2026.
  2. GitHub, Copilot plans and pricing, for "2,000 completions per month", "50 chat requests (including Copilot Edits)", "No credit card required.", and "1 AI credit = $0.01 USD" on the plans that receive credits. Read 13 August 2026.
  3. Base44, Pricing, for the free plan's "25 message credits/mo". Read 13 August 2026.
  4. Linear, AI credits, for "Funds you add expire 12 months after the purchase date" and a $10 minimum ad-hoc top-up. Read 13 August 2026.
  5. Framer, How AI credits and Agent pricing work, for "top-up credits are a one-time purchase and expire at the end of the calendar month in which they're bought" and "any credits you haven't used don't carry over". Read 13 August 2026.
  6. Credit figures and task costs: Clixad's own pricing engine, packages/pricing, over the model catalogue read from OpenRouter on 26 July 2026, recomputed by the test beside this post. The twelve-call task shape is read from the markup of the earlier post that declared it.
  7. Offerwall payouts: completed CPX offers as recorded in Clixad's own ledger, credited between 27 July 2026 and 12 September 2026 and read on 14 September 2026, held as CPX_COMPLETIONS in packages/pricing.