Business Finland
Are OpenAI and Anthropic API fees eligible R&D costs under Business Finland?
Model API spend is one of the least-answered questions in AI R&D budgeting. Here's how to think about eligibility and how to justify it to evaluators.
Short answer: Yes, model API spend can be an eligible cost in a Business Finland R&D budget — but only the portion that’s genuinely tied to the R&D work, not to running the product afterward. API calls made to run experiments, evaluate approaches, or test a hypothesis during the R&D period are budgetable, usually under the “materials and services” or subcontracting-adjacent cost lines. API calls that quietly become your production inference bill are not R&D cost, and evaluators will notice the difference even if the invoice line item looks identical.
This question comes up in almost every AI R&D budget conversation now, because it didn’t exist five years ago. Business Finland’s eligible-cost categories were written for a world where R&D materials meant lab equipment or software licenses, not a metered API bill that scales with usage. Nobody has published a clean answer, so most founders either over-claim (treating all API spend as R&D) or under-claim (leaving legitimate R&D cost out of the budget because they’re not sure it counts).
The distinction that actually matters
Business Finland eligible costs are built around one core question: is this spend necessary to resolve the project’s R&D uncertainty, during the period the uncertainty is being resolved? Applied to API spend, that splits cleanly into two buckets.
Eligible: experimentation and evaluation spend. Running comparative evals across model providers to determine which architecture handles your domain problem. Testing prompting or fine-tuning strategies against a held-out dataset to validate an approach. Running load or accuracy experiments to answer a specific technical question named in your application. This is R&D cost in the same sense a lab consumable is R&D cost — it’s spent to generate knowledge, and it stops (or drops sharply) once the uncertainty is resolved.
Not eligible: production or operational spend. Once the model choice is settled and the API calls are what your live product or internal tool runs on day to day, that’s operating cost, not R&D. It doesn’t matter that the vendor is the same or the endpoint is the same — the purpose has changed.
How to budget it so evaluators believe it
Don’t lump API spend into one undifferentiated line. Break it out against the specific work packages and uncertainties it serves, the way you’d break out any other eligible cost:
- Estimate API usage per experiment or evaluation run, not per month
- Tie each chunk of estimated spend to the uncertainty it’s testing (“evaluating retrieval architectures against domain dataset X”)
- Show the spend tapering as the R&D period ends and, if relevant, production usage begins outside the funded budget
This does two things. It makes the R&D framing legible to an evaluator skimming a budget table, and it protects you at reporting time — if an auditor asks why API costs are in the budget, you have a ready answer that isn’t “we use a lot of AI.”
Where this gets murky
Two situations founders ask about most:
Ongoing evaluation infrastructure that outlives the R&D phase. If you build a continuous eval harness that keeps calling APIs to monitor model quality after the R&D period ends, the harness build itself was R&D; its ongoing operation afterward usually isn’t.
Fine-tuning API costs (training runs, not inference). These sit closer to “eligible” by default, since the training run itself is usually the R&D activity — see the companion piece on fine-tuning below for where that line moves.
FAQ
Do all API providers get treated the same way? Yes — Business Finland doesn’t care which vendor issued the invoice. The test is purpose (R&D experimentation vs operational use), not provider.
Can I budget for API costs I haven’t incurred yet? Yes, budgets are estimates. Base the estimate on a reasonable per-experiment usage projection, and be ready to explain the assumption if asked.
What if my whole product is a thin wrapper around an LLM API? Then very little of your ongoing API spend is R&D-eligible, because there’s likely little R&D uncertainty either — see the piece on what Business Finland doesn’t fund for AI, linked below.
Should I confirm this with Business Finland directly before applying? Yes. This piece describes the reasoning pattern that holds up; exact eligible-cost rules and cost-category names can shift between funding rounds, so confirm current guidance before finalizing a budget.
The one-sentence version
API spend is eligible when it’s the cost of resolving your project’s R&D uncertainty, not the cost of running the finished thing — budget it against specific experiments, not as a blanket monthly line.
Related: Eligible costs in a Business Finland R&D application · Can foundation-model fine-tuning qualify as R&D, or is it “just implementation”? · How to write the R&D uncertainty section of a Business Finland application