# Evanesca — full text

> Evanesca is a private AI agent and chat — a European alternative to ChatGPT. It runs AI models from Mistral AI, the Swiss AI Initiative, Qwen, DeepSeek, Z.ai, OpenAI, Meta and Google in Europe, searches the web, writes documents and creates images, and keeps no copy of your messages. It works in the browser, costs only what you use (credit from €2, no subscription), and shows a passport under every answer: model, operator, place, law, and what was kept.

Source pages: https://evanesca.app (English), https://evanesca.app/fr (French). Third-party policies checked on 2026-09-29; the catalogue and the chat's features on 2026-10-20. Published by FLEECE AI (SASU, SIREN 932 329 188), Chamalières, France.

Status: the chat app opens soon; https://evanesca.app/chat is a waiting page. This site describes how it will work.

## AI that leaves no copy behind.
Evanesca is a private AI agent and chat — a European alternative to ChatGPT. It runs AI models from Mistral AI, the Swiss AI Initiative, Qwen, DeepSeek, Z.ai, OpenAI, Meta and Google in Europe, searches the web, writes documents and creates images, and keeps no copy of your messages. Give it a task — compare three quotes, draft the memo, make the cover image — and it works step by step, with its sources. While web search is off (the default), your messages never leave Europe, and Evanesca forgets them once the answer ends.

Every answer carries a passport: where it was produced, under which law, and what was kept — no content.

## The things you didn't dare ask AI. Now you can.
Most AI chats keep what you send, on servers under US law — sometimes to train their next model. That's why more than one company in four has banned them at some point (Cisco, 2024). Evanesca keeps no copy of what you send, so there is no archive to leak.

What people can now bring to an AI chat without Evanesca keeping a copy: Client contracts (Summarise, compare and redline — without the contract leaving Europe.); Payroll and HR files (Check payslips, draft letters, summarise policies: kept confidential.); Health questions (Ask what you would never type into a US chatbot.); Source code (Debug your product without anyone keeping your code.); Financial results (Analyse figures before they are public — without them being kept.); Board and strategy (Draft the memo about the deal nobody knows about yet.).

## Works like the AI chat you know.
1. Pick a model — or let Auto choose: 17 models, all run in Europe — Auto picks one for each message.
2. Ask, or give it a task: Write, summarise, translate, analyse — or switch on Agent mode: it searches, writes documents and creates images.
3. Check the passport: Under every answer: where it ran, under which law, and what was kept — no content.

## Your messages never leave Europe.
In plain words: every model Evanesca serves runs in Europe — in France at OVHcloud or Scaleway, in Mistral AI's EU region, in Germany at AKI.IO, or in Switzerland at Infomaniak — and the answer comes straight back to you. No model call goes through the US — not as a backup, not when servers are busy. Web search is off by default. If you switch it on, the queries the model writes go to Linkup, a French company — Evanesca has not verified where its servers run. Leave web search off when a message must stay in Europe.

## What changes when you switch.
Same kind of assistant, same kind of answers. The difference is what happens to your words afterwards.

| Question | A typical AI chat | Evanesca |
|---|---|---|
| Where are my messages processed? | In the US or wherever the provider runs — with no option to keep them in Europe | In Europe — France, Germany, Mistral AI's EU region, Switzerland |
| Are they kept? | Until you delete them — often 18 months by default | Not by Evanesca — gone when the answer ends. Some operators keep requests a few weeks to check for abuse (see the FAQ) |
| Are they used to train AI? | Often by default, unless you opt out | Not by Evanesca, nor by its European operators — except Mistral Large 4, a preview Mistral AI may train on |
| Which laws can reach my messages? | EU law — and US law too: the CLOUD Act reaches US providers, even on EU servers | European law where the models run — each at a European operator, and no copy of your messages is kept. The website itself runs on Vercel, a US company, in its Paris region, without keeping a copy |
| How do I pay? | Free with limits, or about €20 a month | Only for what you use — no subscription |
| Which models can I use? | Usually one company's own models | 17 models run in Europe: Mistral, Apertus, Qwen, DeepSeek, Llama… |

"A typical AI chat" means the consumer plans of ChatGPT, Claude, Gemini, Microsoft Copilot and Meta AI, whose policies were read on 2026-09-29:

| Service | Company | Where processed | Kept by default | Trains on chats by default | Read by people |
|---|---|---|---|---|---|
| ChatGPT | OpenAI (United States) | “In the US and around the world”; no EU-only option on consumer plans | Until you delete them, then up to 30 days | Yes, unless you opt out | Yes — staff and service providers |
| Claude | Anthropic (United States) | “Transferred to our servers in the US” | Deleted chats erased within 30 days; up to 5 years if you allow training | Your choice at sign-up | When a chat is flagged for safety review |
| Gemini | Google (United States) | “Servers around the world”; no EU-only option for consumers | 18 months by default; 72 hours even with activity off | Yes, with “Keep Activity” on — the default for adults | Yes — “a subset of chats” |
| Microsoft Copilot | Microsoft (United States) | “In your region, in the United States, and in any other jurisdiction” where it operates | Older app: 18 months by default (the app of 18 Aug 2026 is covered by another page) | Older app: yes, unless you opt out; the app released on 18 Aug 2026 does not say | Older app: yes — “an opt-out of human review is not available” |
| Meta AI | Meta (United States) | Not stated | Deleting “may not delete our copy” | Yes — interactions “will also be used to train” | Yes — review “may be automated or manual (human)” |

### Where your messages are processed
The big consumer chats process conversations in the United States or wherever their providers run, and their consumer plans have no setting to keep them in Europe. OpenAI reserves data residency for eligible business, education and API customers.
- OpenAI: "Content is stored on OpenAI systems and our trusted service providers' systems in the US and around the world." (https://help.openai.com/en/articles/7039943-how-openai-handles-data-in-consumer-services)
- Anthropic: "when you access our Services, your personal data is transferred to our servers in the US, or to other countries outside the European Economic Area ('EEA') and the UK" (https://www.anthropic.com/legal/privacy)
- Microsoft: "Personal data collected by Microsoft may be stored and processed in your region, in the United States, and in any other jurisdiction where Microsoft or its affiliates, subsidiaries, or service providers operate facilities." (https://www.microsoft.com/en-us/privacy/privacystatement)
- Google: "We maintain servers around the world and your information may be processed on servers located outside of the country where you live." (https://policies.google.com/privacy)
- OpenAI: "Eligible API customers and new ChatGPT Enterprise/Edu customers" (https://help.openai.com/en/articles/9903489-data-residency-and-inference-residency-for-chatgpt) — Who can choose where data is stored — consumer plans are not on the list.

### How long they keep your chats
By default, conversations are kept until you delete them — or for 18 months at Google, and in Microsoft's older Copilot app — and copies can remain for weeks after deletion.
- OpenAI: "Regular and archived chats remain saved to your account until you delete them or an applicable workspace retention policy removes them." (https://help.openai.com/en/articles/8983778-chat-and-file-retention-policies-in-chatgpt) — wording re-read on 6 October 2026
- Google: "If you're 18 or over, Keep Activity is on by default." (https://support.google.com/gemini/answer/13278892?hl=en) — The same page: activity older than 18 months is auto-deleted by default.
- Microsoft: "By default, we store conversation activity for 18 months." (https://support.microsoft.com/en-us/microsoft-copilot/privacy-faq-for-microsoft-copilot) — This page describes the Copilot app released before 18 August 2026; the newer app and the web version are covered by another page.
- Anthropic: "Deleted from our back-end storage systems within 30 days" (https://privacy.claude.com/en/articles/10023548-how-long-do-you-store-my-data)
- Meta: "deleting individual messages, entire threads, or even your account may not delete our copy of your personal information" (https://www.facebook.com/policies/other-policies/ais-terms)

### Whether they train on your chats
Most leaders use consumer conversations to improve their models unless you switch it off; Anthropic asks at sign-up.
- OpenAI: "When you use our services for individuals, such as ChatGPT and Codex, we may use your content to train our models." (https://help.openai.com/en/articles/5722486-how-your-data-is-used-to-improve-model-performance)
- Meta: "People's interactions with Meta AI – like questions and queries – will also be used to train and improve our models." (https://about.fb.com/news/2025/04/making-ai-work-harder-for-europeans/)
- Microsoft: "You can opt out of use of your conversation activity for model training at any time" (https://support.microsoft.com/en-us/microsoft-copilot/privacy-faq-for-microsoft-copilot) — This page describes the Copilot app released before 18 August 2026.

### Who can read them
All five leaders say people may review some conversations — and several ask you not to type anything confidential.
- Google: "Please don't enter confidential information that you wouldn't want a reviewer to see or Google to use to improve our services" (https://support.google.com/gemini/answer/13594961?hl=en)
- OpenAI: "Please do not enter sensitive information that you would not want reviewed or used." (https://help.openai.com/en/articles/7039943-how-openai-handles-data-in-consumer-services)
- Microsoft: "an opt-out of human review is not available" (https://support.microsoft.com/en-us/microsoft-copilot/privacy-faq-for-microsoft-copilot) — Copilot app released before 18 August 2026.
- Meta: "this review may be automated or manual (human)" (https://www.facebook.com/policies/other-policies/ais-terms)

### Which law can reach them: the CLOUD Act
A US law, 18 U.S.C. § 2713, lets US authorities order a provider under US jurisdiction to hand over data it controls — even when that data is stored in Europe. The GDPR still applies to those companies in the EU; the point is that a US order can reach the same data.
- 18 U.S.C. § 2713: "within such provider's possession, custody, or control, regardless of whether such communication, record, or other information is located within or outside of the United States" (https://www.law.cornell.edu/uscode/text/18/2713)

### Requests sent outside Europe at busy times
Some services let European requests run outside Europe when demand peaks. Microsoft documents it for Microsoft 365 Copilot, its business product: “flex routing” is on by default for eligible tenants created after 25 March 2026, and administrators can switch it off.
- Microsoft: "If flex routing is enabled, LLM inferencing may occur in the United States, Canada, or Australia during times of peak demand." (https://learn.microsoft.com/en-us/microsoft-365/copilot/copilot-flex-routing)

Privacy-first alternatives also exist:
- Proton Lumo (Proton AG, Switzerland): "After Lumo processes your query and generates a response, the data is erased." (https://proton.me/support/lumo-privacy)
- Kagi Assistant (Kagi Inc., United States): "By default, threads are deleted after 24 hours of inactivity." (https://help.kagi.com/kagi/ai/assistant.html)
- Brave Leo (Brave Software, United States): "We don't store or retain prompts, responses, context, or personal data on our servers." (https://brave.com/privacy/browser/)
- Mistral Vibe (formerly Le Chat) (Mistral AI, France): "Input and output data are used by default to train our artificial intelligence models, unless you opt out of training at any time." (https://help.mistral.ai/en/articles/347617-do-you-use-my-user-data-to-train-your-artificial-intelligence-models)

## What the law asks — and what we do.
The three laws that decide what happens to your words, in plain words. Each rule sits next to what Evanesca does about it.

| Article | The rule | What Evanesca does |
|---|---|---|
| GDPR Art. 5(1)(e) | Keep personal data no longer than the purpose needs. | Prompts and answers are never stored by Evanesca — kept in memory only while the answer is written. Account and billing data: kept only as long as the law requires. |
| GDPR Art. 28 | Processors act under a written contract with sufficient guarantees. | Business workspaces get a DPA to sign and our sub-processor list, with notice before any change. |
| GDPR Art. 44 | Data may leave the EU only under an adequacy decision, safeguards such as standard clauses, or a narrow exception. | Every model runs in the EU or in Switzerland, which has an EU adequacy decision. Evanesca serves no model outside Europe — not even as an option. |
| AI Act Art. 50 | Since 2 August 2026, people must know they are dealing with an AI, and generated content must be marked. | Every answer is labelled as AI output in the app; generated images and Word, PDF and Excel files carry AI-generated metadata; machine-readable marking of answer text is being built. |
| AI Act Art. 53 | The makers of general-purpose models must document them. | We link each model's card, licence and GPAI status from every passport. |
| CLOUD Act § 2713 | Providers under US jurisdiction can be ordered to hand over data they control, wherever it is stored. | Nothing to hand over: no copy of your messages is kept, and every model runs at a European operator. The site itself runs on Vercel, a US company, in its Paris region. |

## Every model shows its checks before you use it.
Six checks, run by us on the route Evanesca uses. A model's passport shows the stamps it has — and the ones it misses or hasn't had yet. We don't hide them.
1. Licence — read at the source, so you know what you may do with the output
2. Provenance — traced to the maker's official release, as declared by the operator
3. Security — injection, jailbreak, leak, exfiltration and harmful-request tests
4. GDPR — a sovereign route: European operator, country and retention documented
5. Benchmarks — quality and speed measured on our own route
6. Live test — a real conversation, end to end, through the route Evanesca uses

## The answer passport
Every answer carries a passport: the model and its maker, the operator and the country where it ran, the law that applies, and what was kept. Its fields: PROCESSED IN, KEPT BY EVANESCA, USED FOR TRAINING, LEFT THE EU, plus the checks the model passed.
- OVHcloud: OVHcloud keeps billing data only.
- Scaleway: Scaleway may keep and read requests that cause errors or look abusive, for up to 14 days.
- A passport describes the route Evanesca gave a request and what the operator states publicly; it is not an audit of the operator's systems.
More: https://evanesca.app/passport

## Models run in Europe (17)
| Model | Maker | Licence | Size | Run by | Where | Checks | Price per M tokens, in / out (VAT incl.) |
|---|---|---|---|---|---|---|---|
| Mistral Large 4 | Mistral AI | — | 1T (49B active) | Mistral AI | EU region of Mistral AI | 4/6 (missing: Licence, Provenance) | €2.513 / €7.725 |
| DeepSeek V4.1 Flash | DeepSeek | MIT | 763B | AKI.IO | Germany (AKI.IO) | 6/6 | €1.008 / €3.024 |
| GLM-5.3 | Zhipu / Z.ai | GLM-5.3 License | 753B | AKI.IO | Germany (AKI.IO) | 6/6 | €3.024 / €9.24 |
| Qwen3.8 27B | Qwen | Apache-2.0 | 27.8B | Scaleway | Paris, France | 6/6 | €1.008 / €5.544 |
| DeepSeek V4 Flash | DeepSeek | MIT | 284B (13B active) | Scaleway | Paris, France | 6/6 | €0.672 / €1.344 |
| Apertus v1.5 70B | Swiss AI Initiative | Apache-2.0 | 72B | Infomaniak | Switzerland (Infomaniak) | 4/6 (missing: Security, Benchmarks) | €1.294 / €4.62 |
| GLM-5.2 | Zhipu / Z.ai | MIT | 753B | Scaleway | Paris, France | 6/6 | €3.024 / €9.24 |
| Qwen3.6 35B | Qwen | Apache-2.0 | 36B (3B active) | Scaleway | Paris, France | 6/6 | €0.42 / €2.52 |
| Gemma 4 26B | Google | Apache-2.0 | 26B (4B active) | AKI.IO | Germany (AKI.IO) | 6/6 | €0.168 / €0.84 |
| Mistral Small 4 | Mistral AI | Apache-2.0 | 119B (6.5B active) | Mistral AI | EU region of Mistral AI | 5/6 (missing: Security) | €0.277 / €1.109 |
| Qwen3.5 397B | Qwen | Apache-2.0 | 397B (17B active) | Scaleway | Paris, France | 6/6 | €1.008 / €6.048 |
| GPT-OSS 120B | OpenAI | Apache-2.0 | 117B (5.1B active) | Scaleway | Paris, France | 5/6 (missing: Security) | €0.252 / €1.008 |
| Qwen3 Coder 30B | Qwen | Apache-2.0 | 30.5B (3.3B active) | Scaleway | Paris, France | 5/6 (missing: Security) | €0.336 / €1.344 |
| Qwen3 235B | Qwen | Apache-2.0 | 235B (22B active) | Scaleway | Paris, France | 5/6 (missing: Security) | €1.26 / €3.78 |
| Mistral Small 3.2 | Mistral AI | Apache-2.0 | 24B | Scaleway | Paris, France | 5/6 (missing: Security) | €0.252 / €0.588 |
| Llama 3.3 70B | Meta | Llama 3.3 Community License | 70B | Scaleway | Paris, France | 5/6 (missing: Security) | €1.512 / €1.512 |
| Pixtral 12B | Mistral AI | Apache-2.0 | 12B | Scaleway | Paris, France | 5/6 (missing: Security) | €0.336 / €0.336 |

DeepSeek V4 Flash runs on Scaleway in Paris and DeepSeek V4.1 Flash at AKI.IO in Germany: prompts go neither to DeepSeek nor to China (https://evanesca.app/models/deepseek-v4-flash).


## Everything you expect from an AI assistant.
- Web search: Answers from today's web, with their sources linked. Behind the scenes: Off by default — a switch in the chat; Only the model's queries go out — to Linkup, a French company; Billed per search, plus the pages read.
- Image generation: Pictures from a description, or portraits from your own photos. Behind the scenes: Made by Infomaniak, in Switzerland; Sent back to your browser, not kept by Evanesca; Billed by the second of generation.
- Documents: PDF, Word, Excel and CSV, read and compared — and written, in Agent mode. Behind the scenes: Text extracted in your browser, sent with your message; Not kept by Evanesca after the answer; Never used for training by Evanesca.
- Photos and scans: Photos, receipts and whiteboards, read and understood. Behind the scenes: Read by a vision model run in Europe; No copy kept by Evanesca; Metadata (GPS, device) removed before upload.
- Charts: Ask for a chart, get a real one — with its data as CSV. Behind the scenes: Drawn in your browser; Data never leaves the thread; Its data downloadable as CSV.
- Code: Highlighting, diffs and live HTML previews, in any language. Behind the scenes: Previews run sandboxed, with no network; No code ever runs on Evanesca's servers; Secrets you paste are never logged by Evanesca.
- Dictation and read-aloud: Speak your question; have the answer read back to you. Behind the scenes: Dictation: Mistral AI, a French company — kept 30 days against abuse; Read-aloud: your device's own voices — nothing leaves it; Dictation billed by the second.
- Memory: Optional: it remembers who you are and what you work on. Behind the scenes: Kept in your browser, readable and erasable; Updated by a small billed call; Off until you switch it on.
- Auto mode: Not sure which model to use? Evanesca picks one, message by message. Behind the scenes: Chooses among the models run in Europe; No extra call, no routing fee; Each answer still names its model.

## Pay only for what you use.
No subscription. You add credit to your account — from €2 — and each answer uses a little of it. You see the exact cost under every answer — usually a fraction of a cent.
- to start: €2 of credit (No subscription)
- your credit: never expires (An invoice each time you add credit)
- Business: Signed Art. 28 DPA, Every model run in Europe, SSO and roles, VAT invoices, Spending limits, Passport exports.

## FAQ
### What is Evanesca — a European alternative to ChatGPT?
Evanesca is a private AI agent and chat — a European alternative to ChatGPT. It runs AI models from Mistral AI, the Swiss AI Initiative, Qwen, DeepSeek, Z.ai, OpenAI, Meta and Google in Europe, searches the web, writes documents and creates images, and keeps no copy of your messages. You use it in the browser, pay per use, and every answer carries a passport: where it ran and under which law.

### What is Agent mode?
You give a task instead of a question — compare offers, draft a report, build a spreadsheet — and the model works step by step: it searches the web, reads your files, writes Word, PDF and Excel documents and creates images. It never runs code, stops at the spending cap you set (€2 by default), and its files stay in your browser. How Agent mode works.

### Does web search send my conversation anywhere?
No. Web search is off by default. If you switch it on, the model writes short queries and picks pages to read; only those queries and addresses go to Linkup, a French company, which fetches the results. Your conversation itself is not sent, though a query can repeat words of your message. Each search and each page read is billed, plus the pages the model reads as input tokens; the exact cost shows under the answer.

### Does Evanesca keep my conversations?
No. Evanesca keeps no copy of your prompts or answers: they stay in memory only while the answer is written, and your history lives in your browser. About your messages, we keep only billing data — tokens and cost. Two exceptions on the operator side: when requests cause errors or look abusive, Scaleway may keep and read their content for up to 14 days; and Mistral AI keeps its API traffic (Mistral Large 4, Mistral Small 4 and dictation) for 30 days to monitor abuse.

### Who can read what I type — Evanesca, the cloud, a government?
Nobody at Evanesca: there is no copy to read. A European operator — OVHcloud, Scaleway, Mistral AI, AKI.IO or Infomaniak — processes your words to write the answer; Scaleway may read requests that cause errors or look abusive, and Mistral AI keeps its API traffic 30 days against abuse. Nobody keeps a copy — not Evanesca, not its host — so a US order under the CLOUD Act has nothing to reach, and the operators are European companies.

### Are my conversations used to train AI?
Never by Evanesca. OVHcloud, Scaleway, AKI.IO and Infomaniak state that they don't train on customer data; Mistral AI lets its customers object to training, and Evanesca's account objects, except for what is sent to Mistral Large 4: it is a preview, and Mistral's commercial terms let it use the data of preview models for training, whatever the objection. About your messages, we keep only billing data — tokens and cost, never content.

### Is Evanesca GDPR-compliant?
Evanesca is built for the GDPR: every model is served in the EU or in Switzerland (EU adequacy decision) by European operators with no US parent, Evanesca never stores your prompts, and Business workspaces get an Article 28 DPA. Your history stays in your browser. No company can certify itself compliant, so we publish what we do, article by article.

### Can I use Evanesca for confidential or work documents?
Yes: every model Evanesca serves runs in Europe, and contracts, payslips, code and unpublished figures are not kept by Evanesca. Web search is off by default — leave it off for them. If the data belongs to your employer or your clients, use a Business workspace: it adds a signed DPA, SSO and spending limits.

### Is DeepSeek safe to use on Evanesca?
Evanesca's DeepSeek models run in Europe: DeepSeek V4 Flash on Scaleway in Paris, and DeepSeek V4.1 Flash at AKI.IO in Germany. Your prompts reach neither DeepSeek nor China. Each model's passport shows the checks it has passed — all six for V4 Flash. Evanesca does not serve DeepSeek V4 Pro: no European operator runs it for Evanesca.

### Is it free? Is there a subscription?
It isn't free, and there is no subscription. You add credit to your account from €2, VAT included — enough for hundreds of everyday questions — and each answer uses a little of it; the credit never expires. The exact cost shows under every answer, usually a fraction of a cent.

### Why are some well-known models missing?
Because Evanesca serves only models run in Europe. Some frontier models — open-weight ones among them — are not run by a European operator for Evanesca, so it does not offer them, not even as an option. More models will join when a European operator runs them.

### What does the AI Act change for me?
For most uses of a chat assistant, the duties are transparency and basic AI literacy for your team. Evanesca tells you that you are talking to an AI, labels its answers and links each model's documentation. Some uses carry more: HR decisions about people are high-risk under the AI Act, with rules applying from December 2027.

## FAQ (français)
### Qu'est-ce qu'Evanesca — une alternative européenne à ChatGPT ?
Evanesca est un agent et un chat IA privés — une alternative européenne à ChatGPT. Elle fait tourner en Europe des modèles d'IA de Mistral AI, la Swiss AI Initiative, Qwen, DeepSeek, Z.ai, OpenAI, Meta et Google, cherche sur le web, rédige des documents et crée des images, et ne garde aucune trace de vos messages. Elle s'utilise dans le navigateur, se paie à l'usage, et chaque réponse porte un passeport : où elle a tourné, sous quelle loi.

### Qu'est-ce que le mode Agent ?
Vous confiez une tâche plutôt qu'une question — comparer des offres, rédiger un rapport, monter un tableau — et le modèle avance pas à pas : il cherche sur le web, lit vos fichiers, rédige des documents Word, PDF et Excel et crée des images. Il n'exécute jamais de code, s'arrête au plafond de dépense que vous fixez (2 € par défaut), et ses fichiers restent dans votre navigateur. Comment fonctionne le mode Agent.

### La recherche web envoie-t-elle ma conversation ?
Non. La recherche web est désactivée par défaut. Si vous l'activez, le modèle rédige de courtes requêtes et choisit des pages à lire ; seules ces requêtes et ces adresses partent chez Linkup, une société française, qui ramène les résultats. Votre conversation elle-même n'est pas envoyée, même si une requête peut reprendre des mots de votre message. Chaque recherche et chaque page lue sont facturées, en plus des pages lues par le modèle, comptées en jetons d'entrée ; le coût exact s'affiche sous la réponse.

### Evanesca garde-t-elle mes conversations ?
Non. Evanesca ne garde aucune trace de vos prompts ni des réponses : ils restent en mémoire le temps d'écrire la réponse, et votre historique reste dans votre navigateur. De vos messages, nous ne gardons que la facturation — jetons et coût. Deux exceptions côté opérateur : quand des requêtes provoquent des erreurs ou paraissent abusives, Scaleway peut en garder et en lire le contenu jusqu'à 14 jours ; et Mistral AI garde 30 jours le trafic de son API (Mistral Large 4, Mistral Small 4 et dictée) pour surveiller les abus.

### Qui peut lire ce que j'écris — Evanesca, le cloud, un gouvernement ?
Personne chez Evanesca : il n'y a pas de copie à lire. Un opérateur européen — OVHcloud, Scaleway, Mistral AI, AKI.IO ou Infomaniak — traite vos mots pour écrire la réponse ; Scaleway peut lire les requêtes en erreur ou jugées abusives, et Mistral AI garde 30 jours le trafic de son API contre les abus. Personne n'en garde de copie — ni Evanesca, ni son hébergeur — : une injonction américaine au titre du CLOUD Act n'a rien à atteindre, et les opérateurs sont des sociétés européennes.

### Mes conversations servent-elles à entraîner l'IA ?
Jamais chez Evanesca. OVHcloud, Scaleway, AKI.IO et Infomaniak déclarent ne pas entraîner de modèles avec les données de leurs clients ; Mistral AI permet à ses clients de s'y opposer, et le compte d'Evanesca s'y oppose, sauf pour ce qui est envoyé à Mistral Large 4 : c'est une préversion, et les conditions commerciales de Mistral l'autorisent à utiliser pour l'entraînement les données des modèles en préversion, quelle que soit l'opposition. De vos messages, nous ne gardons que la facturation — jetons et coût, jamais le contenu.

### Evanesca est-elle conforme au RGPD ?
Evanesca est conçue pour le RGPD : tous les modèles sont servis dans l'UE ou en Suisse (décision d'adéquation) par des opérateurs européens sans maison mère américaine, Evanesca ne stocke jamais vos prompts, et les espaces Business reçoivent un DPA article 28. Votre historique reste dans votre navigateur. Nul ne peut s'autocertifier conforme : nous publions ce que nous faisons, article par article.

### Puis-je y mettre des documents confidentiels ou professionnels ?
Oui : tous les modèles qu'Evanesca sert tournent en Europe, et contrats, fiches de paie, code ou chiffres non publiés ne sont pas conservés par Evanesca. La recherche web est désactivée par défaut — laissez-la coupée pour eux. Si les données appartiennent à votre employeur ou à vos clients, passez par un espace Business : DPA signé, SSO, plafonds de dépenses.

### DeepSeek est-il sûr sur Evanesca ?
Les modèles DeepSeek d'Evanesca tournent en Europe : DeepSeek V4 Flash chez Scaleway à Paris, et DeepSeek V4.1 Flash chez AKI.IO en Allemagne. Vos prompts ne vont ni chez DeepSeek, ni en Chine. Le passeport de chaque modèle montre les contrôles qu'il a passés — les six pour V4 Flash. Evanesca ne sert pas DeepSeek V4 Pro : aucun opérateur européen ne l'exécute pour elle.

### C'est gratuit ? Y a-t-il un abonnement ?
Ce n'est pas gratuit, et il n'y a pas d'abonnement. Vous rechargez dès 2 €, TVA comprise — de quoi poser des centaines de questions courantes — et ne payez que ce que vous utilisez ; le crédit n'expire jamais. Le coût exact s'affiche sous chaque réponse, en général une fraction de centime.

### Pourquoi certains modèles connus manquent-ils ?
Parce qu'Evanesca ne sert que des modèles exécutés en Europe. Certains modèles de pointe — dont des modèles à poids ouverts — ne sont exécutés pour Evanesca par aucun opérateur européen : elle ne les propose donc pas, pas même en option. D'autres modèles rejoindront le catalogue quand un opérateur européen les exécutera.

### Que change l'AI Act pour moi ?
Pour la plupart des usages d'un assistant conversationnel, les obligations sont la transparence et une culture de l'IA de base pour vos équipes. Evanesca vous dit que vous parlez à une IA, signale ses réponses et relie la documentation de chaque modèle. Les décisions RH sur des personnes, elles, sont à haut risque (règles applicables en décembre 2027).

## API
Base URL: https://evanesca.app/api/v1. OpenAI's chat-completions format — not OpenAI's Responses API — so OpenAI's official SDKs work once the base URL is changed; the same chat models also answer in Anthropic's Messages format on POST /messages, for Claude Code and Anthropic's SDKs (their base URL: https://evanesca.app/api; key in x-api-key or as a bearer token). Calls that run a model carry an API key as a bearer token (Authorization header); GET /models needs no key. Every response that reached an operator names that operator and the place in two response headers. The API is not open yet: this describes how it will work.

- GET /models — List models: The catalogue, with no key: chat, embedding and image models, each with its compliance record. No key needed. Reference: https://evanesca.app/docs/api/list-models
- GET /models/{model} — Retrieve a model: One model by its ID: where it runs, its price, its checks, what its operator keeps. No key needed. Reference: https://evanesca.app/docs/api/retrieve-model
- POST /chat/completions — Create a chat completion: Runs a chat model on a conversation: streaming, tool calling, reasoning models. Needs a key. Reference: https://evanesca.app/docs/api/create-chat-completion
- POST /messages — Create a message: The same chat models in Anthropic's Messages format, for Claude Code and Anthropic's SDKs. Needs a key. Reference: https://evanesca.app/docs/api/create-message
- POST /messages/count_tokens — Count message tokens: A free estimate of the input tokens of a Messages request, before you send it. Needs a key. Reference: https://evanesca.app/docs/api/count-message-tokens
- POST /embeddings — Create embeddings: Turns one text or many into vectors, billed per input token. Needs a key. Reference: https://evanesca.app/docs/api/create-embeddings
- POST /images/generations — Create images: Images from a prompt with FLUX.1 [schnell], billed per second of generation. Needs a key. Reference: https://evanesca.app/docs/api/create-image
- POST /images/edits — Create portraits: Portraits from 1 to 4 face photos with PhotoMaker V2, billed per second. Needs a key. Reference: https://evanesca.app/docs/api/create-image-edit

Errors come back as JSON in OpenAI's shape, an error object with a message and a type:
- 400 invalid_request: the body is not valid JSON, a required field is missing, or an image parameter is out of bounds (the param field says which).
- 401 invalid_api_key: no key, a malformed key, or a revoked one.
- 402 insufficient_credit: the credit is at zero or below.
- 404 model_not_found: an unknown ID, a model Evanesca does not serve, or a model of the wrong kind for the endpoint.
- 502 upstream_error: the operator could not be reached, sent an unreadable answer, or turned down Evanesca's own account.
On POST /messages and /messages/count_tokens, the same types come in Anthropic's error envelope ({"type": "error", "error": {type, message}}).

Docs: https://evanesca.app/docs · API reference: https://evanesca.app/docs/api

## Publisher
FLEECE AI SASU, 10 rue Pierre Poisson, 63400 Chamalières, France. SIREN 932 329 188. VAT FR45932329188. Contact: contact@fleeceai.agency