Eden AI vs Requesty

Why

Work the fee arithmetic first, because it is the reason people arrive at this comparison, and it resolves faster than expected. Eden AI marks up nothing per token: provider list prices pass through and the platform fee is 5.5% applied at checkout when you buy credits. Requesty does the opposite: no top-up fee, no minimum spend, and a flat 5% markup on every token, falling to 0% when you route to your own provider keys. Neither charges for seats. Two consequences follow. Requesty's rate is lower on the same consumption, 5% against 5.5%, and Requesty charges only for what you actually use, while Eden AI's fee lands on the balance rather than the burn — and Eden AI's terms state that payments are non-refundable, so a fee paid on credits you never spend is gone. Set against that, Eden AI's fee is charged once per top-up regardless of how the credits are later consumed, so both converge on roughly a twentieth of spend, and both go to zero if you bring your own keys: Eden AI publishes no BYOK surcharge at all, though with BYOK the cost it returns becomes an estimate rather than a charge.

So the fee shape is nearly a wash, and the real split is what each vendor does with your prompts. Eden AI records zero data retention as the documented default rather than an enterprise mode: retention is 0 days for request and response content, content is simply not stored unless you switch on the per-project Log Retention toggle, and it states in its terms, privacy policy and security page that customer data is not used to train models. Requesty defaults to 30 days of logs holding full request and response bodies. Organisation-wide zero retention is available and enforced server-side, but it must be requested in writing and it also disables Requesty's own caching, so the compliance win costs you the cache. And the free plan is a different animal: models labelled Training Permitted retain prompts and outputs and train on them, and those models are free-plan only. Requesty's paid plans are absolute on training; the free tier is not.

Residency and attestation split the same way, which is unusual for a pair this closely priced. Eden AI is a French company headquartered in Lyon with EU-hosted infrastructure and a dedicated EU host that refuses to route outside the EU, plus per-request region pinning, all self-serve; it records SOC 2 as claimed, though no report type is named, and GDPR compliance. Requesty is a UK company with an EU endpoint in Frankfurt where logging, caching and analytics all stay in region — a stronger statement about what happens to the data than a routing restriction alone — but its SOC 2 Type II is recorded as in progress and expected in Q3 2026, not attested. If your procurement needs a named report today, neither record gives you one cleanly, and you should ask both for the document rather than trusting either page.

Catalogue and control surface then push in opposite directions. Requesty publishes 600+ models with a live models call returning 686 entries, against Eden AI's headline 500 and a live catalogue of 1,010 endpoints collapsing to 488 distinct routable LLM names; on providers it inverts, with 33 distinct LLM providers counted from Eden AI's own catalogue against Requesty's 20 stated and only 6 listed as connectable for BYOK. Operationally Requesty is the more complete router: semantic caching, PII scanners with Disabled, Report and Mask actions, guardrails, an MCP Gateway, retries from 0 to 10 with exponential backoff, and a vendor claim of roughly 16 ms of added overhead against about 55 ms for managed OpenRouter and 124 ms for self-hosted LiteLLM, published without methodology. Eden AI has no semantic cache at all, though its exact-match cache is on by default and a hit costs nothing, which is a real saving on embeddings, moderation and other repetitive calls and is exactly where the vendor recommends it. It has no MCP surface, no timeout, retry or circuit-breaker controls, a fallback list capped at three entries, and it publishes no latency figure of any kind. Its guardrails are asserted on the marketing security page and appear nowhere in the documentation index as a feature, parameter or example — treat them as unproven and plan to call the moderation endpoint yourself.