Eden AI

What is Eden AI?

Eden AI is a unified AI API gateway provided by a French company; it helps developers access multiple third-party AI providers using a single account, a set of keys, and a set of interfaces. It addresses issues such as repeated connections, scattered billing, model switching, failover mechanisms, and usage monitoring.

Eden AI is not a provider of underlying large models; rather, it acts as a technical intermediary and governance layer. The quality of the models, their availability, as well as data processing and output are still influenced by the supplier chosen.

Current product range

The current V3 documentation states that access is available to more than 500 models from over 50 suppliers, covering both general language models and specialized AI capabilities. The list of models is continuously updated, and some newly added pages use different criteria for classification; therefore, it is necessary to consult the current list of models when making a selection.

Unified interfaces for V3 category 2

V3 organizes its key capabilities into a chat interface and a Universal AI interface. The chat interface is compatible with the common OpenAI Chat Completions request format, while the expert model interface handles tasks such as OCR, text processing, image analysis, audio processing, and translation.

All requests are authenticated using a Bearer API key. This key should be stored only on the server side or in a controlled key management system; it must not be included in web pages, mobile application packages, or public repositories.

Model switching, intelligent routing, and fallback

Developers can explicitly select suppliers and models, or they can use intelligent routing to let Eden AI choose models based on quality, cost, latency, and the supplier’s performance. The pool of available options can be restricted to limit the automatic selection process and prevent selection of models that do not meet internal policies.

• Explicit models are suitable for scenarios that require fixed behavior, a test baseline, and repeatability.
• Intelligent routing is suitable for dynamically balancing quality, cost, and latency.
• The candidate pool may only include models that have passed legal, security, and quality reviews.
• Fallback attempts the alternative models in sequence when the main model fails, is subject to rate limiting, or is unavailable.
• The fallback to an expert model requires that the primary supplier and the backup supplier provide support for the same functions.
• The output structure, content policies, and billing units of different models may still vary.

Automatic fallback improves availability, but it may alter the language style, accuracy, region, and costs. Production systems should keep track of the supplier, model, costs, and fallback path that were ultimately used.

Unified billing and cost monitoring

Eden AI provides a unified view of balances, invoices, and usage for different suppliers; each API response includes a field showing the cost in dollars. Teams can create multiple API keys to separate the costs related to testing, production, customers, or specific projects.

• The model catalog displays the supplier prices, which are updated as the suppliers make changes.
• The response cost field helps establish cost metrics for individual requests, users, and functions.
• Automatic top-up allows you to purchase more credits when the balance falls below a certain threshold.
• It is possible to set a monthly limit for automatic top-ups, thereby reducing the risk of unauthorized charges.
• Underlying suppliers may use different units such as tokens, images, minutes, pages, or requests.
• Caching, retries, and fallback options can all affect the actual costs, and these factors should be taken into account during performance testing.

Develop the integration process

1. Register an account and create an API key for the development environment in the console.
2. Query the model and feature catalog to filter by capabilities, region, price, and input constraints.
3. Choose a compatible interface for chat tasks, and Universal AI for tasks such as OCR.
4. Configure the base address and Bearer authentication on the server side, so that the keys are not exposed on the client side.
5. First, send a minimal request using a single inexpensive model to check the status, output, and cost fields.
6. Logic for handling timeouts, supplier errors, rate limiting, and insufficient balance is written separately for each case.
7. Add tested fallback options, and set thresholds for the quality and cost associated with using those fallback solutions.
8. Create separate keys and budget monitoring for production, testing, and different customers.
9. Define the permitted suppliers, models, and processing regions based on data classification settings.
10. Complete the review of DPA, terms with third-party suppliers, as well as logging and deletion procedures prior to going live.

Compatible integration with OpenAI

The chat interface uses an OpenAI-compatible request format; existing tools can generally be integrated by replacing the base address, API key, and model name. Compatibility refers to similar call structures, though it is not guaranteed that every provider will support exactly the same parameters, tool calls, or workflow patterns.

• First, lock in the client version and send a minimal chat request.
• Test system messages, structured output, tool calls, images, and streaming responses.
• There is logic for downgrading or removing parameter settings that are not supported by a certain model.
• Do not assume that the model name, context length, and price remain constant forever.
• Logs network request errors and the internal supplier failure status in the responses.
• Pattern verification, content review, and manual recheck are carried out on the key results.

Prices and packages

The self-service version does not add a surcharge to the supplier’s unit price; instead, it charges a 5.5% platform fee at the time of settlement based on the purchase amount. Payments are processed through Stripe. The unit prices and billing units vary depending on the supplier, and the final cost should be determined according to the model catalog, the cost field displayed for each response, and the settlement page.

The terms indicate that the service can be accessed on a pay-as-you-go basis, through prepaid payment, or via subscription; unless otherwise specified, payments are generally non-refundable. Before enabling automatic top-up, it is necessary to set thresholds, monthly limits, and billing alerts.

Speed limits and capacity

The current price comparison table indicates 10 requests per second for the self-service version; the detailed specifications on the same page state that the default value is 7 requests per second, with the possibility to increase this to 15 requests per second. The figures provided may vary, so it is necessary to refer to the information displayed in the account, the details given by technical support, or the terms of the contract.

• Lower-level suppliers may impose stricter independent speed limits than those set for Eden AI accounts.
• The limits on concurrency, tokens per minute, file size, and asynchronous tasks may vary depending on the feature.
• Fallback cannot guarantee bypassing the common capacity issues of all suppliers.
• The enterprise version allows for negotiated higher speed limits and bulk discounts.
• Stress testing should cover peak loads, fallback paths, timeouts, retries, and cost limits.

Data streams and default retention

Eden AI’s privacy policy and DPA state that prompts, outputs, and customer content are not stored by default; they are processed temporarily solely for the purpose of routing requests, with only the limited metadata necessary for billing, security, and operations being retained. The optional logging feature changes this default setting.

• Account data can include name, email address, company information, and payment details.
• Using metadata can include time, request volume, model, token, and costs.
• Customer data may include text, files, images, audio, and model outputs.
• The default setting of retaining no content does not mean that all metadata are deleted immediately.
• The file reuse feature requires a persistent identifier; it is necessary to determine the specific actions for saving and deleting files before using it.
• Eden AI claims that it does not sell customer data nor uses customer content to train its models.

The security page states that prompts, uploaded files, and output are retained by default for zero periods of time; it also specifies that the policy of \"deletion within 24 hours\" applies in such cases. Enterprises should define clearly in contracts, DPA agreements, and actual configurations what is meant by \"zero retention,\" as well as any exceptions and procedures for deleting relevant evidence.

The underlying suppliers still need to be evaluated separately.

Eden AI sends requests to the third-party model providers chosen by the user, and DPA specifies that these providers may act as sub-processors for handling data. The default retention and security policies of the Eden AI gateway do not automatically cover all the actions of each underlying provider.

• Check whether the supplier maintains prompt logs, output logs, or abuse monitoring logs.
• Identify the data processing location and the required cross-border transmission mechanisms.
• Check whether the supplier uses data to improve or train models.
• Understand the content policy, copyright terms, and restricted usage scenarios.
• Smart routing and Fallback are enabled only in the pool of candidates that have passed the review.
• Sensitive data should be minimized, anonymized, and subject to access controls before being fed into the model.

Security, Compliance, and Governance

Eden AI’s security page indicates that it holds SOC 2 and ISO 27001 certifications, and it provides documents related to GDPR as well as DPA agreements. The enterprise features include control over endpoints in different regions, management of suppliers and models, input protection, output auditing, policy-based routing, and centralized visibility.

Terms, Output, and Responsibilities

The terms specify that Eden AI does not own or operate the underlying models, nor does it control their availability, performance, outputs, or data processing practices. Customers are required to comply with the applicable terms of their chosen supplier and are responsible for their own data, applications, and intended uses.

• AI outputs may be incorrect, incomplete, harmful, or infringe on third-party rights.
• Medical, legal, financial, and security decisions cannot rely solely on model outputs.
• Customers must have a legitimate basis to process and transmit personal data.
• The product and service license does not grant access to the weights of the underlying models, the training data, or the supplier’s source code.
• Model switching may change the output rights and acceptable scope of use.
• High-risk applications should have mechanisms for manual approval, audit logging, and revocation.

GitHub, SDKs, and open-source status

Making Skills, examples, documentation, or integration code available publicly does not mean that the Eden AI gateway and the underlying models are open source. Since the V3 chat interface is compatible with the OpenAI format, many teams can use compatible clients without having to rely on a specific SDK.

Suitable for users and typical use cases

• Development teams that need to test quality, latency, and cost across multiple LLMs.
• SaaS products that simultaneously utilize OCR, speech, translation, image, and text models.
• Enterprise platform teams that seek to unify keys, billing, monitoring, and supplier management.
• Production applications that require fault recovery and the ability to replace suppliers.
• European or regulated projects that require EU endpoints, DPA, and regional policies.
• Developers who wish to create agents and multimodal workflows through the same interface.

Advantages and main limitations

Main advantages

• A single interface provides access to a large number of language models as well as expert AI capabilities.
• Unified billing, cost fields, and multiple keys facilitate cost attribution.
• Smart routing, candidate pools, and Fallback enhance switching capabilities and fault tolerance.
• The OpenAI-compatible format reduces the cost of migrating existing chat applications.
• It offers SOC 2, ISO 27001, DPA, as well as regional and zero-retention controls.

Main limitations

• In addition to the amount used by the supplier, a 5.5% platform fee must also be paid.
• The retention, regions, terms, and speed limits of the underlying suppliers still need to be examined separately.
• On public pages, the default rate limit varies between 7 and 10 requests per second.
• The model catalog, parameter options, and prices may change over time as provided by the supplier.
• Intelligent routing and fallback can introduce differences in quality, cost, and compliance.
• The core network disables its source; the public repository contains mainly documents, examples, and integrations.

Frequently Asked Questions

Does Eden AI offer its own large models?

No. It is a unified API gateway that routes requests to the third-party providers chosen by the users, and it handles authentication, billing, monitoring, and governance in a centralized manner.

How many models does Eden AI support?

The current V3 documentation covers more than 500 models and over 50 suppliers. The catalog is subject to change, so it is necessary to check the specific models using the real-time model list.

How is billing done?

The self-service version consumes the prepaid credit based on the prices published by the underlying suppliers; a 5.5% platform fee is charged additionally when purchasing credits. Each response includes a cost field in dollars.

Is there a permanently free quota?

The current pricing is based on the pay-as-you-go quota; no fixed, permanent free quota is indicated on the price page. Any promotional credits or trial periods available after registration are specified on the account page.

Is automatic model selection supported?

Support is available; smart routing can be used along with restrictions on the candidate models. In a production environment, it is still necessary to keep track of the final model chosen, the associated costs, and any fallback scenarios.

Is there an automatic switch-over in case the supplier fails?

A Fallback list can be specified in the request; in case the primary model fails or throttling occurs, alternative models are tried sequentially. The alternative models must possess the necessary capabilities.

Are the prompt and output saved by default?

The privacy policy and DPA state that, by default, customer prompts and outputs are not stored; only the necessary metadata is retained. Optional logging and the handling by underlying suppliers require further examination.

Does Eden AI have SOC 2 and ISO 27001 certifications?

The current security page indicates that it has obtained SOC 2 and ISO 27001 certifications, and it offers support related to DPA and GDPR. The specific audit reports and scope of application are available upon request during the procurement review.

Is Eden AI open source?

The core gateway is not a product with its source code made publicly available. The official team has provided Skill, documentation, Cookbooks, and some integration repositories, but not the core platform or the weights of the underlying models.

Can the payment be refunded?

The terms state that, unless otherwise specified, payments are generally non-refundable. Before purchasing a large amount of credits or enabling automatic top-ups, it is necessary to set up budgets, alerts, and approval processes first.