There are four proven ways to make money with AI models: sell a service built on someone else’s model (the fastest path, no code required), build and sell a product that wraps an API, train or fine-tune a model and sell access to it, or sell data, prompts and training material to people who do. Service work pays first and most reliably. Product and model sales scale further but take months to reach income. The economics below are current as of October 2026.
Which path pays soonest?
Service work. A freelancer who uses an AI model to deliver copy, video edits, research summaries or voiceovers invoices a client this month. Clients pay for the outcome, not the tool, so your rate is set by the deliverable’s value rather than by the model’s cost.
From a production company’s standpoint, the margin is in the gap between what a model costs per job and what the job bills for. A rough-cut transcript, shot log and first-pass edit script for a 60-minute interview costs well under a dollar in tokens on a mid-tier model and bills as several hours of assistant editor time. The client is paying for judgment, turnaround and accountability, which the model does not supply on its own.
What does it actually cost to run a model?
Hosted models bill per million tokens. Anthropic’s published pricing shows the spread between tiers:
| Model | Input per 1M tokens | Output per 1M tokens |
|---|---|---|
| Claude Haiku 5.5 | $0.10 | $0.50 |
| Claude Sonnet 5.5 | $2.00 | $10.00 |
| Claude Opus 5.5 | $4.00 | $20.00 |
Batch processing cuts both figures in half, and cached prompts cost a fraction of base input. For a product, this means unit cost depends almost entirely on which tier you route each task to. Classification and extraction belong on the cheapest tier. Anything a customer reads verbatim deserves a stronger model.
Image, audio and video models are priced per second of GPU time or per generation. Replicate’s pricing page lists per-second hardware rates, and most open models there cost a few cents per image. Self-hosting on Hugging Face Inference Endpoints starts at under a dollar per hour for small GPUs and climbs quickly for anything that needs an 80 GB card.
How do you build a product on an API without losing money?
Price per outcome, not per seat, until you know your cost per task. A monthly flat fee invites your heaviest users to consume all your margin. Charge per generated asset, per minute of processed footage, or per report, and keep a measured ratio of revenue to token spend. Three to five times is a sustainable floor once support and hosting are counted.
Narrow vertical tools win. A generic chatbot competes with the model vendor’s own app. A tool that turns a real estate listing into a 30-second vertical video script, caption set and shot list competes with nobody in particular and solves a problem a specific buyer already pays for.
Can you sell a model you trained yourself?
Yes, three ways. Sell inference access through a marketplace or your own endpoint. Sell the weights under a commercial license. Or sell fine-tuning as a service, where a client supplies brand voice samples or product footage and you deliver a tuned model they control. The last option is the most practical for a small studio because it stacks on service revenue you are already earning.
Fine-tuning a small open model on a few thousand examples costs tens to low hundreds of dollars in compute. The real cost is assembling and cleaning the training set, which is usually days of human labor.
What about selling data, prompts and training material?
Model builders pay for licensed footage, voice recordings and labeled datasets. A production company sitting on years of released b-roll and signed talent agreements has a licensable asset, provided the releases cover machine learning use. Older releases rarely do, so check before listing anything. Prompt marketplaces exist but pay little. Treat them as marketing rather than income.
What are the legal and tax rules you cannot skip?
Three apply to almost every path.
- Copyright. The U.S. Copyright Office’s AI guidance holds that purely machine-generated output is not protectable, while works with meaningful human authorship are. If your product’s value depends on clients owning the output, your workflow needs documented human creative control.
- Disclosure. The FTC rule on consumer reviews and testimonials bans fake AI-generated reviews and undisclosed paid endorsements. Selling review generation or synthetic testimonials is not a business, it is a liability.
- Taxes. Freelance and product income is self-employment income. The IRS self-employed tax center covers quarterly estimated payments and deductible expenses, which include API bills and GPU rentals.
Where should a beginner start this week?
Pick one deliverable you can already sell, cut its production time with a model, and bill the same price. Track token cost per job for a month. That number tells you whether a product version is viable and what to charge for it. Everything else follows from having paying customers before you have code.
