What free testing looks like at EU inference providers
You have a product that runs on one inference provider and a shortlist of others. Before anyone signs, you want the new provider to answer your own prompts. One buyer asked us in a sales call: can we test your platform for free before we fully commit?
The test is worth running now. Epoch AI found that since January 2026, the most capable open-weight models have trailed the best closed models by about four months. It measures this on its Epoch Capabilities Index, a single score that combines many capability tests. Its data is dated 29 May 2026. Open models are now close enough that testing them on your own prompts is worth the effort. An open-weight model is one whose trained model files are published, so several providers can serve the same model.
You will meet three kinds of free or low-cost test. They pay for different things.
| Offer | What it is | What it tests |
|---|---|---|
| Sign-up credit | A small balance added once to a new account | The shared API, on your own prompts |
| Trial key | An API key with free, limited access for testing | The shared API, within the key's limits |
| Proof of concept (PoC) | A time-boxed test of one workload, with success criteria agreed before it starts | A setup planned for your workload |
Lyceum is a sovereign European AI inference provider for open-source models, meaning built and run in Europe. Our serverless models run in Paris and Finland. Locations for dedicated endpoints and GPU virtual machines (VMs) are agreed per contract. We don't keep your prompts or outputs, and we never use them for training. With serverless inference, you call a shared model through one OpenAI-compatible API and pay per token, with no base fee. A token is the unit of text a model reads and writes, and prices are quoted per million tokens.
How far a $20 sign-up credit goes on a real eval
An eval is a fixed set of your own prompts, each with a check on the answer. It is the test that tells you whether a provider works for you. Take a hypothetical eval of 500 prompts, each with 2,000 input tokens and 500 output tokens. Input tokens are what you send. Output tokens are what the model writes back.
The cost of one run follows from the two prices on the model's page:
cost of one run = (input tokens / 1,000,000 x input price)
+ (output tokens / 1,000,000 x output price)Here is that formula for two models from our models page, at their list prices in US dollars per million tokens. In the table, 1.00 means 500 prompts x 2,000 input tokens = 1.0 million tokens. 0.25 means 500 x 500 output tokens = 0.25 million. Runs are rounded down to whole runs.
| Model | 1,000,000 input tokens | 250,000 output tokens | One run | Full runs on $20 |
|---|---|---|---|---|
| GLM-5.3 | 1.00 x $1.40 = $1.40 | 0.25 x $4.40 = $1.10 | $2.50 | 8 (20 / 2.50) |
| Kimi K3 | 1.00 x $3.00 = $3.00 | 0.25 x $15.00 = $3.75 | $6.75 | 2 (20 / 6.75 = 2.96) |
Calculated example: the eval is hypothetical, and the prices are our list prices, read on 9 November 2026.
On GLM-5.3, the 250,000 output tokens cost $1.10. That is almost as much as the 1,000,000 input tokens at $1.40. Output is priced higher per token, so long answers use up a credit fastest. On Kimi K3, the same eval costs $6.75, and the grant runs it twice.
Real tests usually cost more than this example. It leaves out retries, latency under load and the prompts you spend while you build the eval. Some models also write reasoning text before the answer, which adds output tokens. Count your own tokens from your current provider's usage data before you compare offers.
Trial credit or structured PoC: which test you need
A credit tests the shared API. If you will call serverless models in production, that is the right test. You run your eval on the same models, endpoint and billing you would use later. When the credit runs out, you keep the account and pay as you go, with no commitment.
A PoC tests capacity planned for you. A dedicated endpoint runs one model on GPUs reserved for your workload. Its setup is agreed before launch, so the test needs those GPUs set up for you. A small credit cannot pay for reserved GPUs, so a time-boxed, paid test with written goals fits better. Our PoCs for dedicated inference are paid, and we agree the success criteria with you first. Afterwards, we review the results with you against those criteria. If they aren't met, you're under no obligation.
What goes into the PoC itself is a topic of its own. Our guide on what a two-week PoC with a new inference provider should include covers exit criteria and measurements.
What to check before a trial key touches real data
A free tier is still a contract. Before you send anything beyond test data, check five things:
- Where requests run during the trial, not only on the paid plan. GDPR Chapter V is the chapter on transfers. Under it, personal data sent outside the European Economic Area must meet its transfer rules.
- Whether prompts and outputs are stored, or used for training.
- Whether unused credit expires.
- Whether the paid plan after the trial has a monthly minimum.
- Whether you can cap what a test key spends.
Terms differ on the last three. T-Systems' LLM Hub lets you create a trial key that unlocks all its models. Its paid plans start at a minimum token spend of €1,000 per month. OpenRouter's terms let it expire unused credits 365 days after purchase. Replicate's purchased credit is valid for 1 year. We checked these three pages on 9 November 2026.
On our side, our credits don't expire. There is no monthly minimum. We don't store your prompts. Contracts still come first when real customer data is involved. A data processing agreement (DPA) sets how a provider handles your data. Our checklist for DPA, NDA and customer data before an LLM PoC lists what to sign before that point.
How to run a capped test on our API
A good free test cannot cost more than you planned. Three steps set that up on our API.
- Claim the grant. Open the dashboard and verify a card to claim the sign-up grant. The card check is a Stripe SetupIntent, which saves a payment method without creating a charge.
- Cap the test key. A spend limit is a cap on what one API key may spend on serverless inference. An owner or admin sets it. A monthly limit starts over each month, while a fixed budget stays used up until an owner raises or resets it. Once the key reaches its limit, new requests get HTTP
402with the error codeapi_key_spend_limit_reached. The API keys docs show both settings. - Watch the balance. Charges come off your organisation's balance in real time, so you can stop a long eval early. One call shows free and purchased credits and the total used:
curl https://api.lyceum.technology/api/v2/external/billing/credits/details \
-H "Authorization: Bearer <your-token>"Set the fixed budget to the size of the grant, and the eval stops when the free credit is gone. To size longer tests first, see how to estimate serverless inference costs before you commit.
The decision then follows from the result. If your eval passes on serverless models, keep the account and pay per token as traffic varies. If you need reserved capacity, ask us for a dedicated PoC with the criteria written down first. Your next step: open the dashboard, claim the grant, put a fixed spend limit on a test key and run your eval.
Frequently asked questions
Can I test Lyceum for free before committing, for example with trial credits or a proof of concept?
For serverless inference, you pay as you go with no commitment, so you can start small and scale when you're ready. For dedicated inference, we run time-boxed proofs of concept with success criteria we agree on before you commit. The customer pays for these proofs of concept. They usually run for a few days, up to a week, depending on the workload.
Can several people test on one account?
Yes. Credits live on the organisation, so every member's requests bill against one balance. An owner or admin can give each tester a separate key with its own spend limit.