Pricing

You pay for one thing: the input tokens the engine reads. $0.042 per million input tokens. Output is free. Memory writes are free.

Jepela is not an LLM. It answers typed questions with probabilities and never writes text, so there is no output to pay for.

Get a key

What you pay for

• Input tokens: the state, your questions and their options, and the memory lines recalled for the subject.
• The engine reads the state once per question, so a request with three questions reads it three times.
• usage.input_tokens in every answer is the engine's own count, never an estimate. usage.output_tokens is always 0.
• Recalled memory lines count as input when a decision reads them.
• "robust": true and "windows": true read more, and what they read is billed.
• A choice with more than 20 options is matched by default and billed for the state and memory it read, not its options.
• A cache hit is billed like the first call.

On a 77-option choice (600 bank messages, measured 2026-09-27), matching read 23 tokens where the engine reading every option read 812.

What is free

• Output: always 0 tokens.
• Memory writes: remember and rules.
• "memory": {"compare": true}: the same decision without the memory, to see what it changed.
• "memory": {"explain": true} and "memory": {"placebo": true}: the passes that show what each memory line did.
• A request that fails upstream (502) or in the gateway (500).
• Fine-tuning, during the preview. A fine-tuned model's decisions are billed like any other.

Worked examples

Measured on 2026-09-24 through https://api.jepela.com with jepela-english, on a 74-character support message. The token counts are the engine's own. The cost is arithmetic at $0.042 per million input tokens, before VAT.

Your own requests show their count in usage.input_tokens.

Memory instead of resending history

With memory, you write what you know about a subject once, and each decision reads only the lines it recalls. In the cost cookbook (60 constructed decisions about 20 customers, run 2026-09-26), sending the whole history with every request took 496 tokens per decision. Naming the subject and letting memory recall the lines took 139. The same 60 decisions were used to choose how recall works, so this shows what the tokens cost, not how often memory is right.

Prepaid credit

• Jepela is prepaid. You add credit by card or by invoice.
• Card payments go through our payment provider's checkout page. The provider resells the credit to you and adds VAT there.
• Every priced request is one line in your ledger and is charged against your credit at once.
• GET /v1/usage shows credit, spending, balance, requests and the last twenty ledger lines. The totals and the balance always agree.
• At zero balance, priced requests answer 402 until credit is added. GET routes still work, so you can see your account, and removing data (forget, delete and the other delete routes) still works and never takes the balance below zero.

For business customers

Jepela is for business customers only. Prices exclude VAT and other taxes, which are added where the law requires. For card payments, VAT is handled at checkout.

Rate limits are per account, counted over all its keys. Over the limit, the API answers 429 with a Retry-After header.

Compared with Jev

TypeSafe lists Jev at the same price: $0.042 per million input tokens, output free (TypeSafe models page, read 2026-09-23). With Jepela, memory writes are free too. See Decision APIs compared.

Get a key

Accounts opened on 2026-10-15. Create an account and choose your region: the United States is live, and Europe is planned. Once you confirm your email address, a card payment approves the account at once; otherwise Jepela approves it. Then you create your keys in the console and add credit.

The details, including what each request option reads, are in Pricing in the documentation.