If your AI tool pays for every model call, a credit you give away costs you only the inference behind it, while the user values it at your retail price. This playbook turns that gap into distribution: an invite-only programme that pays your heaviest users in credits for building with the tool and publishing the results.
The idea and why it works
Blanket free credits go to everyone, including people who were never going to pay. This programme rewards one behaviour: power users who build with your tool and put that work in public. Raw use earns a little. Published work with a visible "made with" credit earns more. Published work that brings a sign-up earns the most. More building and publishing means more credits, which keeps members on your tool and gives you showcase content, distribution and product signal.
The cost asymmetry is what makes it profitable. Say a credit is one generation on a mid-tier model. As of 22 Oct 2026, Anthropic lists Claude Sonnet 5.5 at $2 per million input tokens and $10 per million output tokens. An illustrative generation of 4,000 input and 1,500 output tokens costs $0.008 + $0.015 = $0.023, or about $0.025 with a 10% retry buffer. Sell credits at an illustrative $0.10 and each granted credit is worth $0.10 to the member and costs you $0.025.
Who it's for
Founders of AI tools with metered usage and shareable output (image, video, writing, voice, code, agents), a free or entry tier, a visible group of heavy users who already post results, and no ad budget. It fits poorly where output is private, because members have nothing to publish.
What you need
Per-user usage data: active days, generations, projects.
Referral links per member and a "made with" badge or caption.
A credits ledger (schema below) and a public rules page.
Two to three hours a week to review posts and approve credits.
An inference budget. With the illustrative numbers, a 15-person pilot is capped at $2,250 worst case.
Step-by-step
Price one credit. Average a week of real token usage per action, multiply by your provider's current price and add a retry buffer. That is k, your cost per credit.
Invite on usage signals. Illustrative gate: 12+ active days in the last 30, 200+ generations, at least one public post. Start with free and entry-tier users. Credits given to someone already on your top plan replace revenue you would have earned anyway.
Set earning rules and tiers. Illustrative: 1 credit back per 10 used (capped at 100 a month); 50 credits per published piece with "made with" attribution that reaches 200 views after 7 days; 150 credits per verified sign-up through the member's link. Tiers set the monthly ceiling: Starter 300, Builder 800, Partner 2,000. Members move up on monthly results.
Decide where it ends. Lifetime cap of $600 retail value (6,000 credits) and a six-month term, then graduation to paid with an optional loyalty discount. Earned credits expire after 60 days, cannot be cashed and are spent after any paid credits.
Build the ledger and anti-farming checks. Hold publish and sign-up credits as pending until the threshold is met. Cap approvals at, say, three published pieces a day.
Require disclosure. The FTC says people who receive "free products or other perks with the expectation that you'll promote" a product are covered by its rules. It adds: "Even an incentive with no financial value might affect the weight or credibility of an endorsement." Businesses rewarding customer content should "tell them in advance that they should disclose what they received". Starting a post with "Ad:" or "#ad" would likely be effective, the FTC says. Never tie credits to a positive opinion.
Run a 60-day pilot with 10–20 members. Weeks 1–2: invite, onboard, record baselines. Weeks 3–6: approve credits weekly and run one 20-minute call or three-question survey per member. Weeks 7–8: compare members with similar non-members and cost with revenue, then expand, adjust or stop.
Treat graduation as an offer. The logic of the 30-day upsell window applies: pitch the paid plan while the member is at peak use, before their credits lapse.
Template
Invite email
Subject: You're one of our heaviest builders. Want credits for it?
Hi {first_name}, you've made {n_projects} projects with {product} this month, and some are public. We're inviting a small group into a six-month credits programme. You earn credits for using {product}, more for publishing work with a "made with {product}" credit, and the most when someone signs up through your link. Credits are capped at ${cap} in value, expire after 60 days and can't be cashed out. Posts made for credits need an "Ad:" or "#ad" label. Rules: {rules_url}. In return, we'd value one short call or survey a month. Reply "in" and I'll set you up this week. {founder_name}
Programme rules
Invitation only; six months or the lifetime cap, whichever comes first.
Published work earns only after reaching the stated views or sign-ups. Daily caps apply.
Every post made in exchange for credits must be labelled.
Credits expire after 60 days, cannot be cashed, transferred or sold, and are used after paid credits.
Spam, fake accounts, self-referrals or bought views end membership and void pending credits.
Ledger schema
credit_ledger(entry_id, member_id, tier,
event_type, -- usage_rebate | publish_verified | signup_attributed | spend | expiry | adjustment
source_url, ref_code, credits, retail_value, inference_cost_est,
status, -- pending | approved | rejected
created_at, approved_at, expires_at)
members(member_id, invited_at, term_end, lifetime_cap_retail,
retail_granted_to_date, disclosure_ack_at, graduated_at)The maths
Illustrative inputs: $0.10 retail per credit, $0.025 inference per credit, six-month term, $600 lifetime cap.
Tier | Monthly ceiling | Retail value | Your cost | Max cost per member |
|---|---|---|---|---|
Starter | 300 credits | $30 | $7.50 | $45 over 6 months |
Builder | 800 credits | $80 | $20 | $120 over 6 months |
Partner | 2,000 credits | $200 | $50 | $150 (cap hit in month 3) |
Breakeven (illustrative). A $20-a-month plan includes 200 credits costing you $5, leaving $15 a month. Over an assumed 10-month lifetime, a paying customer contributes about $150, the most a capped member can cost. So a capped member breaks even with one paying customer: themselves on graduation, or someone they brought in. Worst case for 15 members is $2,250, or 15 paying customers. If five graduate, the other 10 must come from sign-ups attributed to published work.
Fill-in version
k = (avg input tokens × input price + avg output tokens × output price) ÷ 1,000,000 × 1.1
Programme cost = credits redeemed × k + your hours × your hourly rate
V = (monthly price − inference included per month) × expected months
Breakeven: (graduates + paying sign-ups from published work) × V ≥ programme cost + revenue lost from members who would have paid anyway
Metrics to watch
Credits redeemed cost as % of revenue, monthly, against a ceiling you set before launch.
Graduation-to-paid rate at each member's term end.
Sign-ups attributed to published work, and how many pay.
Retention of members versus similar non-members, matched on pre-invite usage.
Roadmap signal: tag what members build and where they hit limits. With permission, their inputs make good labelled cases for automated evals and prompt tuning.
Pitfalls
Credit-hunters join for credits and stop building. The usage gate and view thresholds filter most out.
Low-quality spam posts. Pay on verified views and sign-ups, apply daily caps, sample-review weekly.
Margin creep. A pricier model or longer outputs raise k quietly. Recalculate monthly from token logs.
Members never graduating. Without a hard cap and end date, the programme becomes a permanent free plan.
Variations
Template bounty: credits for public templates other users adopt, measured by installs.
Cheaper-model credits that run only on a lower-cost model, cutting k. As of 22 Oct 2026, Anthropic lists Claude Haiku 5.5 at $0.10 input and $0.50 output per million tokens for prompts up to 100,000 tokens.