Pricing
Output is free. You pay for input tokens.
Dex never writes text, so nothing is billed on the way out. You pay per million input tokens: the state once, plus your questions. Prepaid in euros. No subscription, no minimum spend beyond a first top-up.
Sign-up and API keys are open, and test keys are free. Paid top-ups open soon.
Per million input tokens
EUR 0.05per million input tokens, at every volume
- 015 micro-cents per billable token. Every charge is an exact whole number.
- 02Output tokens are free. The fixed prompt text around your content is free.
- 03The state is billed once, however many questions share it.
- 04Prices exclude VAT.
What a call is billed forExample from the API contract
What calls cost
| Call | Billable tokens | Charge |
|---|---|---|
| One check on a short message | 150 | EUR 0.0000075 |
| The Dutch support example in the docs | 273 | EUR 0.00001365 |
| A typical 450-token call | 450 | EUR 0.0000225 |
| A 1,000-token state with 8 questions of 60 tokens | 1,480 | EUR 0.000074 |
One million calls of 450 tokens cost EUR 22.50.
Cost and speed
Cost and speed per model.
Median time per workflow against cost per 1,000 workflows, both from our recorded runs. Lower and further left is better. The runs are on the home page.
Cost and speed per modelmean over 4 workflows, 2026-09-25
| # | Model | Median time (s) | Cost per 1,000 (EUR) | Source |
|---|---|---|---|---|
| 1 | Dex, GPU compute | 0.103 | 0.0463 | our recorded runs, 4 workflows |
| 2 | Dex, round trip | 0.130 | 0.0463 | our recorded runs, 4 workflows |
| 3 | GPT-5.4 mini | 0.859 | 1.0327 | our recorded runs, 4 workflows |
| 4 | GPT-5.3 chat | 1.141 | 2.6086 | our recorded runs, 4 workflows |
| 5 | GPT-4.1 | 1.234 | 2.4764 | our recorded runs, 4 workflows |
| 6 | GPT-4.1 mini | 1.475 | 0.4960 | our recorded runs, 4 workflows |
| 7 | Claude Haiku 4.5 | 1.554 | 1.8837 | our recorded runs, 4 workflows |
| 8 | Claude Sonnet 5 | 2.114 | 4.8454 | our recorded runs, 4 workflows |
| 9 | Claude Opus 5.5 | 2.670 | 9.7278 | our recorded runs, 4 workflows |
| 10 | Claude Fable 5.1 | 7.052 | 29.3534 | our recorded runs, 4 workflows |
| 11 | Jev, TypeSafe's published figure (different workload) | 0.400 | 0.3435 | TypeSafe, workflow evals, read 25 Sept 2026 |
Each point is a contender's mean of its per-workflow medians. Dex is shown twice: round trip over the network, and GPU compute alone. Costs are tokens times list price; Dex at its list price, output free. USD converted at 0.858664 EUR per USD (Azure rate, read 2026-09-25).
How you pay
- 01Prepaid balance in euros. Top up EUR 10 to EUR 2,500 at a time, with cards, iDEAL, Bancontact, SEPA Direct Debit.
- 02New accounts can top up at most EUR 500 in their first 30 days.
- 03Each call is charged when it is accepted, and refunded in full if it then fails on our side.
- 04Balances do not expire. Unused balance is refunded on request when you close the account.
- 05VAT: reverse charge for EU businesses with a valid VAT id. Dutch VAT for Dutch customers.
Free test keys
Test keys start with dex_test_. They run on the GPU path, are never billed, and share a free quota of 250,000 tokens per account per day.
Comparison
Dex, the EU alternative to Jev
Jev is a decision model from TypeSafe AI. Like Dex, it answers typed questions about text or JSON with probabilities. This comparison uses only TypeSafe's own public pages. Every row links to its source and says when we read it.
| Topic | Dex | Jev, as TypeSafe's pages state it | Source |
|---|---|---|---|
| Where requests are processedDex differs | In the EU only. Azure West Europe in the Netherlands, Dex's own inference nodes on dedicated hardware operated by thinQit in the Netherlands, and Azure OpenAI in the EU data zone for the fallback path. | TypeSafe's subprocessor list names Amazon Web Services, Modal, Nebius and CoreWeave for request processing, each with the location USA. | TypeSafe Trust Center, Subprocessorsread 25 Sept 2026Subprocessors table, cloud and AI infrastructure rows |
| Request content kept by defaultDex differs | None, for every account. Content is stored only if the account owner turns on content logging. | TypeSafe's legal page offers zero data retention (ZDR) to enterprise customers.One exception on the Dex side. On the fallback path, until Microsoft approves modified abuse monitoring, Azure OpenAI may keep prompts its classifiers flag for up to 30 days inside the EU data zone. Send fallback set to never to stay on our GPU. | TypeSafe docs, Legalread 25 Sept 2026Data retention paragraph |
| Training on customer dataSame on both | Never. Customer content is not used to train, tune or calibrate any model. | TypeSafe's models page says Jev is not trained on customer requests or responses. | TypeSafe docs, Modelsread 25 Sept 2026Data paragraph |
| Same input, same answerNot comparable yet | Designed so that an identical request to the same GPU model version returns byte-identical answers. Fallback answers are excluded and labelled as such.Measured: Published at launch | TypeSafe's consistency cookbook reports 90.8% raw label agreement over 15 repeated calls on one post with 8 questions, sampled on 11 Sep 2026, and says its threshold policy does not make the model deterministic.Not like for like. Ours is a design guarantee that we will verify in production; theirs is one measured experiment. | TypeSafe docs, self-consistency cookbook for choicesread 25 Sept 2026Results table and the note under the threshold policy |
| Abstaining when unsureDex differs | Built in. Set min_confidence on a question and an answer below it comes back with abstained set to true. | TypeSafe's confidence page leaves thresholding on confidence to the caller's own code. | TypeSafe docs, Confidenceread 25 Sept 2026Using confidence section |
| LanguagesDex differs | English and Dutch quality measured and published separately, as agreement with reference labels on our own test set.Measured: EN 84.9%, NL 79.6% | TypeSafe's models page names English as the primary training language, where accuracy is best, and says other languages are handled but not equally well. | TypeSafe docs, Modelsread 25 Sept 2026Languages paragraph |
| List price per million input tokensJev ahead | EUR 0.05, flat at every volume. | USD 0.042, on TypeSafe's models page.Different currencies. We do not convert. Both prices exclude VAT where it applies. | TypeSafe docs, Modelsread 25 Sept 2026Pricing section |
| Output tokensSame on both | Free. | Free, per TypeSafe's models page. | TypeSafe docs, Modelsread 25 Sept 2026Pricing section |
| Largest requestJev ahead | 32,768 billable tokens per request, of which the state can use up to 16,384. | TypeSafe's models page gives 64k tokens per request, of which 32k for the state plus the longest question.Each count uses its own tokenizer, so the numbers are close but not identical in text length. | TypeSafe docs, Modelsread 25 Sept 2026Limits section |
| Published measurements todayJev ahead | Agreement and calibration error of our current model version on our own English and Dutch test set, and recorded races of our prototype against general LLMs on the home page. Production latency and determinism measurements are not published yet. | TypeSafe publishes accuracy, cost and latency on its workflow evals page, and a repeatability experiment in its docs.Our race numbers come from a prototype on Dex's own inference node. The cost and speed chart on the pricing page shows TypeSafe's published figure beside them, marked as a different workload; it is not a like-for-like comparison. | TypeSafe, workflow evalsread 25 Sept 2026Whole page |
Where Jev is ahead today
- 01List price per million input tokens. USD 0.042, on TypeSafe's models page.
- 02Largest request. TypeSafe's models page gives 64k tokens per request, of which 32k for the state plus the longest question.
- 03Published measurements today. TypeSafe publishes accuracy, cost and latency on its workflow evals page, and a repeatability experiment in its docs.
How we made this comparison
Nobody working on Dex has used Jev or TypeSafe's service. We read TypeSafe's public pages and cite them with the date we read them, and we read them again before every change to this comparison.
Every row is a fact from a public page, with the page and the date we read it. Where we have not measured something yet, the Dex cell says so. Where the two are not comparable, the row says why.
- TypeSafe docs, Models read 25 Sept 2026
- TypeSafe docs, Legal read 25 Sept 2026
- TypeSafe docs, Confidence read 25 Sept 2026
- TypeSafe docs, self-consistency cookbook for choices read 25 Sept 2026
- TypeSafe, workflow evals read 25 Sept 2026
- TypeSafe Trust Center, Subprocessors read 25 Sept 2026