Compare gateways

Compare AI gateways

Experiential beside OpenRouter, Kong, LiteLLM, Portkey, Cloudflare, Vercel and Helicone, feature by feature with sources.

At a glance

This page puts Experiential next to seven other AI gateways, one row per capability. We checked every competitor cell against that vendor's own docs or pricing page on 2026-11-19. "Not documented" means we could not find it there, not that it is missing.

ExperientialOpenRouterKong AI GatewayLiteLLMPortkeyCloudflare AI GatewayVercel AI GatewayHelicone
HostingHostedHostedKonnect hosted or self-hostedSelf-hosted, open sourceHosted, hybrid, or open-source gatewayHostedHostedHosted or open source
OpenAI Chat and ResponsesBothBothBothBothBothChat (compat), Responses via OpenAI pass-throughBothChat
Anthropic MessagesNative, every modelYesNative format, Anthropic onlyYesYes, any providerPass-through to AnthropicYesNot documented
Gemini generateContentNative, every modelNot documentedNative format, Gemini and Vertex onlyYesNot documentedPass-through to GoogleNot documentedNot documented
EmbeddingsYesYesYesYesYesPass-throughYesNot documented
BatchesYesYesYesYesYesNot documentedBatch jobs mentioned in pricingNot documented
Price on tokensList price, no markup5.5% fee on credit purchases (Standard)No model resale; plan and request feesFree (open source); Enterprise customPer request: Pro $49 a month for 100K, then $9 per 100K5% fee on Unified Billing creditsList price, no markup0% markup on credits
Bring your own keyNo markup; first paid key needs ProNo fee to $25K a month, then 5%Your keys onlyYour keys onlyYour keysStored keysNo markup; $0.20 per 1K if fallback is offYes
FailoverOrdered waterfall; slow lanes demotedProvider and model fallbacksLoad balancing with fallback (Enterprise plugin)Fallbacks and retriesFallbacks, retries, circuit breakersRetries, model fallbackProvider and model fallbacksAutomatic failover
Zero data retentionPro: org switch or per request; verdict on every responseOrg setting and per requestNot documentedNot documentedNot documentedUnified Billing only, not BYOKPer request free; team-wide $0.10 per 1KNot documented
Spend limitsPer-key caps, budgets, tier rate limitsPer-key limits, guardrail budgetsToken budgets per consumer groupPer key, team and userPer key and providerSpend limit per gatewayPer team, project, key, memberCustom rate limits
Content guardrailsStandard pack or your own classifier, per keyModel, provider and moderation policiesPrompt guards, PII sanitizer, cloud safetyContent filters, PII masking50+ checksLlama Guard, DLPNot documentedNot documented
Response cacheNone (provider prompt caching passed through)Not documentedSemantic cache (Enterprise plugin)Exact and semanticSimple and semanticYesNone (prompt caching only)Yes
Models980 listed, all callable600+17 providers, your keys100+ providers3,000+ models, your keys14 providers on the compat endpointNot documented100+
SSONot yet availableEnterpriseEnterpriseEnterpriseEnterpriseNot documentedNot documentedEnterprise (SAML)

Where Experiential is better

  1. Four wire formats on one hosted key. OpenAI Chat, Responses, native Anthropic Messages and native Gemini generateContent all reach every catalog model, plus embeddings and batches. Among the hosted gateways here, only Experiential documents all four. LiteLLM matches it, but you host it. See Models and the catalog.
  2. Zero data retention you can verify per response. One org switch, or provider: {"zdr": true} on a single request (OpenRouter's field shape). Unknown providers count as not-ZDR. It applies to every lane, your own keys included, on Pro with no per-request fee. Every completion carries x-gateway-zdr and x-gateway-provider, and the list of zero-retention routes is public. Cloudflare's ZDR covers Unified Billing only, and Vercel charges for team-wide ZDR. See Data controls.
  3. No-training routing on every plan. require_no_training needs no paid plan and applies to your own keys too. See Data controls.
  4. Slow lanes get demoted, not just failed ones. A provider that answers slowly moves down the waterfall on its own, then recovers on its own. Failover never cuts you off to protect a cache hit. See Routing and failover.
  5. No markup on tokens, either lane. Credits pay list price. Your own key is billed by the provider, with no per-request fee and no volume threshold (a first paid-provider key needs Pro). OpenRouter charges 5% on BYOK above $25K a month, and Vercel charges $0.20 per 1K BYOK requests with fallback off. See Plans and pricing.
  6. A drop-in OpenRouter surface. It keeps the same /api/v1 paths, vendor/slug model ids, :free spellings, and /generation, /credits and /key reads. Moving off OpenRouter is a base URL and key change. See Compare: OpenRouter.
  7. Cost in the response. Finished Chat, Responses and Messages replies carry usage.cost and the provider that served them, so billing a request needs no second lookup.
  8. Guardrails with your own classifier. Per key, run the standard PII, secret and injection pack, or call your own HTTP endpoint. Prompts and completions are never stored by the guardrail. See Content guardrails.
curl -si https://api.experientiallabs.ai/v1/chat/completions \
  -H "Authorization: Bearer YOUR_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{"model": "gpt-6-luna", "provider": {"zdr": true},
       "messages": [{"role": "user", "content": "Hello"}]}' \
  | grep -iE "^x-gateway-|^x-request-id|\"cost\""

When another gateway may fit better

  • You must self-host or run air-gapped. LiteLLM, Kong, Portkey's hybrid or open-source gateway, and Helicone's open-source build run in your network. Experiential is hosted.
  • You already run Kong for API management. Keep it, and point its ai-proxy upstream at Experiential (see Kong).
  • You need a response cache. Portkey, LiteLLM, Cloudflare and Helicone cache whole responses, and Kong offers a semantic cache. We pass provider prompt caching through but do not cache responses.
  • You need SSO today. OpenRouter, Kong, LiteLLM, Portkey and Helicone offer SAML or OIDC on Enterprise plans. Ours is not yet available.
  • Your app lives on Vercel or Cloudflare Workers. Their gateways plug into OIDC and Worker bindings with no key to manage.

Sources, checked 2026-11-19