Compare gateways
Experiential vs LiteLLM
LiteLLM is a proxy you host on your own keys, while Experiential is the same kind of surface run for you, and it can be LiteLLM's upstream.
Summary
LiteLLM is a self-hosted, open-source proxy for 100+ providers. It has virtual keys, budgets, fallbacks, caching and guardrails, and SSO and audit logs on Enterprise. You run it, and you hold every provider account. Experiential serves the same wire formats as a hosted service, with model access on credits. You can keep LiteLLM in front and point its models at Experiential.
model_list:
- model_name: fast
litellm_params:
model: openai/gpt-6-luna
api_key: os.environ/OPENAI_API_KEY
- model_name: smart
litellm_params:
model: anthropic/claude-sonnet-5
api_key: os.environ/ANTHROPIC_API_KEYSide by side
| Experiential | LiteLLM | |
|---|---|---|
| Hosting | Hosted | Self-hosted (open source); Enterprise self-hosted or air-gapped |
| APIs | Chat, Responses, Messages, Gemini generateContent, embeddings, batches | All of those, plus audio, images, realtime, rerank and more |
| Model access | 980 listed models on credits, or your keys | 100+ providers, your keys |
| Pricing | List price on tokens, no markup | Free open source; Enterprise custom |
| Failover | Ordered waterfall; slow lanes demoted | Fallbacks, retries, load balancing |
| Zero data retention | Pro: org switch, per request, verdict per response | Not documented (you host the proxy; provider terms are yours) |
| Spend limits | Per-key caps, budgets, tier rate limits | Per key, team, user and tag |
| Content guardrails | Standard pack or your own classifier | Content filters, PII masking |
| Response cache | None | Exact (Redis, S3, GCS, memory) and semantic |
| SSO | Not yet available | Enterprise |
Where Experiential is better
- Nothing to run. No proxy, database or Redis to operate, patch and scale.
- No provider accounts to open, and catalog prices are known, so cost is exact per request (
usage.cost). LiteLLM needs custom pricing for models it does not know. - Zero-retention and no-training routing as one switch, enforced on every lane. See Data controls.
- Failover that also demotes slow providers. See Routing and failover.
When LiteLLM may fit better
You must self-host or run air-gapped, need a response cache, or need endpoints we do not serve. Keep LiteLLM and add Experiential as an upstream.
Migration notes
Two options. Keep LiteLLM and change each model_list entry (shown above): clients and virtual keys stay the same. Or remove LiteLLM and point clients at https://api.experientiallabs.ai/v1.
| What | LiteLLM | Experiential |
|---|---|---|
| Base URL | http://localhost:4000 (your proxy) | https://api.experientiallabs.ai/v1 |
| Auth | sk- virtual key or master key | Authorization: Bearer YOUR_API_KEY |
| Model ids | provider/model (anthropic/claude-sonnet-5) | Plain slug (claude-sonnet-5); inside LiteLLM, openai/<slug> |
| Virtual keys, budgets | /key/generate with max_budget, rpm_limit | One Experiential key per team, with per-key caps |
Request metadata | Tags, trace ids | X-Explabs-Tags, X-Explabs-Session-Id headers |
| Spend reads | /key/info, spend logs | usage.cost on each reply, the Cost API |
Switch with an agent
Paste this into your coding agent. It inventories the project first and asks before anything it cannot map.
POST/v1/chat/completionsimport litellm
litellm.completion(
model="anthropic/claude-sonnet-5",
messages=[{"role": "user", "content": "Hello"}],
)Sources, checked 2026-11-19
- LiteLLM proxy: 100+ providers, virtual keys, budgets, fallbacks, guardrails, logging, Enterprise SSO.
- Supported endpoints: Chat, Responses, Messages,
generateContent, embeddings, batches, files and more. - Caching: Redis, semantic, S3, GCS, local.
- LiteLLM Enterprise: custom pricing, SSO, SCIM, audit logs, self-hosted and air-gapped.
- Experiential: /llms.txt, GET /api/models.