Pricing

Pay only for what you use

No monthly fees. No commitments. No minimum spend. Point your AI workflows at one API and pay only for the AI Runs you consume.

An AI Run is one call to the API

It bills as 600 tokens, and input tokens count half — about what an ordinary email and its reply take up. A long prompt does not penalise you, no call ever bills below 0.125 runs, and a call that fails bills nothing at all.

AI Runs = (input_tokens × 0.5 + output_tokens) / 600

What would you be paying here?

Put in the AI bill you already pay and the model it goes to. The same work, priced against this catalogue.

What you spend on AI today€ a monthOn which modelWhat you would run here

Today€500.00

Here€129.53

74% less, every month

That bill is about 129,534 AI Runs a month.

An estimate, not a quote. It assumes the same work on both sides — an average call with a 600-token prompt and a 300-token answer, which is exactly one AI Run — and it does not include the one-off 5% platform fee at top-up. Closed-model prices are the providers' published list prices, checked on August 20, 2026; ours are the ones in the table below.

Prefer your exact number? Audit your usage exportExport the usage CSV from your provider's dashboard (OpenAI: Usage → Export), drop it here, and every call is re-billed with the real formula — no average-call assumption.

Your usage export (CSV)

…or choose the file

The file is read in your browser and sent nowhere. This page makes no request with it.

Every call is re-billed as max(0.125, (input × 0.5 + output) / 600) AI Runs at auto's price from the catalogue below. List prices are the providers' published ones, checked on August 20, 2026; dollar costs convert at the same dated rate. On aggregated exports the per-call floor applies to averages, so this is an honest, floor-aware estimate — not an invoice.

Choose your level of control

Same API, same balance, same invoice. The only question is who picks the model.

Auto

recommended

You define the goal. We pick the model for each request.

€1.00per 1,000 AI Runs — flat, whoever answers

A model chosen per request: trivial jobs go to the fast one, everything else to the best all-rounder
One flat price whatever it takes to answer — our routing bill, not yours
Failover to a second provider when the first one is down, at no extra cost
Shadow sampling, if you switch it on: one call in a hundred is answered by a cheaper model too and the two are scored against each other — billed to us, not to you
Full transparency: every response carries an X-Resolved-Model header naming the model that actually answered

One price, not three quality tiers. Guessing which tier a request deserves is the work you came here to stop doing. Today the router defaults to gpt-oss-120b, and we move it when a better model appears.

Manual

You choose the model. For builders who already know what they need.

from€0.25per 1,000 AI Runs

Pin a model to a key, or name it per request
One API for models from several providers
One balance and one invoice for all of them
Change model without touching a single workflow
A test bench that runs your real cases against the whole catalogue before you switch

Every price in the catalogue is published below.

The whole catalogue, at provider cost

What you pay per call is the provider's own price for those tokens, passed through to the cent — no per-call markup, no volume tiers, no negotiated rates. Prices change through the catalogue changelog.

Chat models
Model	What it is for	Provider cost (what you pay)
auto recommended	We pick the model per request: simple tasks go to the fastest one, everything else to the best all-rounder. Flat price, whoever answers. · today it resolves to gpt-oss-120b	€1.00€ per 1,000 AI Runs€0.15 in · €0.60 out per million tokens
gpt-oss-20b	The cheapest model here. Reasons briefly before answering; plenty for classification, tagging and short replies at volume.	€0.25€ per 1,000 AI Runs€0.04 in · €0.15 out per million tokens
mistral-small	Fast and cheap. Ideal for classification, data extraction and short replies.	€0.75€ per 1,000 AI Runs€0.15 in · €0.35 out per million tokens
qwen3.5-9b	Small, recent, and with genuinely long context. The budget pick when the prompt is large.	€0.40€ per 1,000 AI Runs€0.10 in · €0.15 out per million tokens
gpt-oss-120b	Balanced quality and price. Our default choice.	€1.00€ per 1,000 AI Runs€0.15 in · €0.60 out per million tokens
gemma-4-26b	Understands images and writes fluently. A lot of model for the price.	€1.25€ per 1,000 AI Runs€0.25 in · €0.50 out per million tokens
deepseek-v4	Thinks before it answers: slow to start, and the thinking is billed even though you never see it. The best here for code and multi-step problems.	€1.75€ per 1,000 AI Runs€0.40 in · €0.80 out per million tokens
qwen3.6-35b	A nimble current-generation reasoner. Thinks before it answers, and the thinking is billed — as with every reasoner.	€2.25€ per 1,000 AI Runs€0.25 in · €1.50 out per million tokens
qwen3.6-27b	A dense 27B of Qwen's 3.6 generation. Reasons by default and the thinking is billed, as with every reasoner.	€3.50€ per 1,000 AI Runs€0.40 in · €2.70 out per million tokens
qwen3.8-27b	Qwen's newest generation in a dense 27B. Reasons by default and the thinking is billed; on open-ended planning it can spend the whole budget thinking, so give it room or ask for less.	€4.50€ per 1,000 AI Runs€0.60 in · €3.30 out per million tokens
llama-3.3-70b	Natural writing and reliable instruction following.	€2.25€ per 1,000 AI Runs€0.90 in · €0.90 out per million tokens
qwen3-235b	The most capable model here for complex, multilingual work.	€3.75€ per 1,000 AI Runs€0.75 in · €2.25 out per million tokens
qwen3.5-397b	Qwen's big reasoner. Takes images and video as input, and competes with the closed models on hard tasks.	€4.75€ per 1,000 AI Runs€0.60 in · €3.60 out per million tokens
glm-5.2	Open frontier for code and agents. The big model you pick when the result matters more than the price.	€10.00€ per 1,000 AI Runs€1.80 in · €5.50 out per million tokens
mistral-medium	Mistral's big one: strong multilingual, image understanding, reasoning. European end to end.	€11.00€ per 1,000 AI Runs€1.50 in · €7.50 out per million tokens

Embeddings, transcription and speech
Model	What it is for	Price
embed	Turns text into vectors for semantic search.	€0.35per million tokens
embed-large	Higher-quality vectors and input texts four times longer than "embed". Same price.	€0.35per million tokens
whisper	Transcribes audio to text in over 50 languages.	€10.00per 1,000 minutes
whisper-turbo	Fast transcription at a third of whisper's price. For volume: calls, meetings, voicemail.	€3.00per 1,000 minutes
speak-es	Turns text into Spanish speech. Female or male voice.	€5.00per million characters
speak-en	Turns text into English speech. Female or male voice.	€5.00per million characters
speak-de	Turns text into German speech. One male voice.	€5.00per million characters
speak-it	Turns text into Italian speech. Female or male voice.	€5.00per million characters

Every price change, with its date

What the task costs, not what a token costs

The cheapest price per token is not the cheapest invoice. Three things sit between the two, and this is where each one lands here.

Retries nobody asked for
Ask for JSON, get something that will not parse, and somebody has to ask again. We ask again for you and bill the attempt we delivered — never both.
The fallback when a provider is down
Failing over to the second backend is another call at somebody else's price. Here it is the same price: you are billed for the alias you asked for, whoever ends up answering it.
A unit that moves under you
A token count depends on what the model decides to write that day. An AI Run is 600 tokens with input at half weight, and no call bills below 0.125 of one — so the same task costs the same this month and next.

Enroutia Credits: top up, then pay provider cost per call

Four packs, from 20 €. A 5% platform fee is charged once, when you top up — 20 € of credit is 21 € before tax — and after that each call debits exactly what the provider charges. No monthly fee, no per-call markup, nothing renews unless you switch auto top-up on yourself.

Credit	Platform fee (5%)	You pay (before tax)	Best for
€20.00	€1.00	€21.00	Trying it out
€50.00	€2.50	€52.50	Personal projects
€100.00	€5.00	€105.00	Builders and small agencies
€500.00	€25.00	€525.00	Production workloads

The 5% fee is charged once, at top-up, and never per call. Load 20 €, pay 21 €, and every call after that costs what the provider charges us.
Credit you bought never expires. Only the welcome credit does, after 30 days, and we warn you before it goes.
Run out and calls return a clear error with a link to top up. No automatic charges, no overage, no debt.

Three things that will not happen to you

No per-call markup
The only thing we charge is a 5% platform fee, once, when you top up. Every call after that debits the provider's real cost — the number in the table above, not a number above it.
Your balance does not expire
Credit you bought stays there until you spend it. Only the welcome credit expires, and we warn you first.
Your price is the one from your last top-up
If prices rise, your balance keeps paying the price you saw when you bought it, until your next top-up. If they fall, you get it right away.

Enroutia Credits since 6 September 2026: packs of 20, 50, 100 and 500 €, one 5% fee at top-up, provider cost per call.

Two questions about credit

What happens when I run out of credit?: Calls return a clear error (402) with a link to top up. No automatic charges and no debt.
Does credit expire?: Welcome credit expires after 30 days. Credit you buy does not.

Always know what is happening

Every AI Run leaves a trail, and all of it is yours to read:

Which model answered, on every response, in the X-Resolved-Model header — plus a Warning header on the calls that were served under a different name than the one they asked for
Cost, tokens and AI Runs by day, by model and by key, in the panel
What routing saved you this month, against sending the same work to the dearest model you used
How many calls asked for something that is not in the catalogue, and which names they used
Latency and time to first token, call by call
The whole event history, downloadable: model, tokens, price, latency, which provider served it and when

No black box. And none of it includes what you sent: prompts and answers live in memory for the length of the request and are gone.

How our privacy works

And control it before it happens

A monthly cap per key, so a runaway workflow stops instead of emptying the balance
Spend alerts by day, week or month, by email or webhook
One key per project or per client, each with its own cap and its own report
Auto top-up if you want it, with a daily ceiling and a monthly cap. Off by default
A monthly budget for the whole workspace, on top of the caps per key

Replace several AI bills with one API

Before

n8n → OpenAI API
n8n → Anthropic API
n8n → Mistral API

After

n8n → Enroutia API → the model that fits the request

One integration. One invoice. One balance. And in your workflow, this is the only line that changes:

OPENAI_BASE_URL=https://api.enroutia.com/v1

European privacy, by architecture

“We cannot read your prompts” is a claim about architecture, not about our intentions. This page explains what makes it literally true, and where its limits are.

What we cannot see

The content of your requests and responses is never written to any disk: not to the database, not to logs, not to backups. The web server's access logs are switched off, so there is not even a record of which IP address requested which path. Application logs carry only errors and security events, with the IP passed through a hash with a daily salt.

Who else is involved

Your prompts are processed on the inference provider's servers, in the European Union, with no retention and no training. Saying less than this would be a lie: somebody has to run the model, and that somebody sees the text for as long as it takes to answer.

The honest limit

Real end-to-end encryption does not exist for inference: the model has to read the prompt to answer it. The serious version of that promise — confidential computing on GPU with attestation — is on the roadmap, with our own infrastructure. We will not promise it before we can demonstrate it.

And three more things worth knowing

Third parties in your browser

This site loads no external fonts, scripts or CDNs. The two exceptions are the anti-abuse captcha, only on signup and the comparator, and the payment provider's script, only on the payment screen. The analytics we use are cookieless, which is why there is no banner.

Confidential Mode

Any key can turn it on. With it, that key never reads or writes the response cache: your content does not spend even an hour in shared memory.

The only thing we store is what you hand us

Your traffic prompts are not stored anywhere, and that does not change. The first exception is the test bench: if you press “Save as test”, that specific prompt is stored encrypted so it can be re-run against other models and show you whether a cheaper one does the same job. You choose it, prompt by prompt, there is a limit of 10, and deleting it really deletes it — the prompt and its answers.

How our privacy works

For agencies running several clients

Group your keys by end client, cap each client monthly, add your markup and produce a statement with your own logo. It is in the panel from the first day, on this same pay-as-you-go price list — there is no agency tier to buy.

Running something bigger than that, or need what is not here yet? Write to us and we will answer as people, not as a form.

Sign-ups are closed while we run this for ourselves

Leave an address and we will write once, when there is a key to hand you. Nothing else.