LiteLLM vs Standard Compute
LiteLLM is free software that you host, secure and connect to every provider yourself, while Standard Compute is one hosted endpoint with the compute included and a router that picks the cost-efficient model per request, which suits teams that would rather not run a proxy.
Side by side
| LiteLLM | Standard Compute | |
|---|---|---|
| Pricing | Free and MIT-licensed; Enterprise license through sales; you pay each provider | Flat monthly plan: $19, $39, $89 or $249 with $20 to $269 of compute included; Business from $499 |
| Who picks the model | Tiers you assign; cost strategy within one model group | Router picks per request (harder steps to stronger models, routine steps to efficient ones), or pin any model ID |
| Budget control | Hard max_budget per virtual key, user and team | Plan budget per key; new requests get HTTP 402 when it is used, no overage |
| Claude Code | Native on your proxy | Native Messages API |
| Codex | Documented (Responses API) | Native Responses API, any model in the catalog |
| Europe / US rules | Wherever you host it, with the providers you choose | Opt-in Europe or US rules on every plan, per key, team or company |
| Teams | Virtual keys, users and teams | Company and agency workspaces, a plan per key |
| Best for | Platform teams that want to self-host | Coding agents and agent fleets that need one fixed monthly cost |
LiteLLM facts from its own pricing and docs pages. Standard Compute is our product.
Where they differ
- LiteLLM is a proxy you deploy, upgrade and secure; in March 2026 two of its PyPI releases shipped a credential stealer. Standard Compute is hosted.
- With LiteLLM you keep an account and a bill with each provider. Standard Compute is one plan with the compute included.
- LiteLLM's auto router sends prompts to tiers you assign by hand, and its cost strategy chooses between deployments of one model. Standard Compute picks the model per request.
When LiteLLM is the better fit
LiteLLM is the better pick for a platform team that wants no third party in the data path, the widest provider coverage, and full control through configuration.
Switch from LiteLLM in 2 minutes
Standard Compute speaks the same APIs, so switching is a base URL, a key and a model name. Your prompts, tools and streaming stay as they are.
| Setting | LiteLLM | Standard Compute |
|---|---|---|
| OpenAI-compatible base URL | http://localhost:4000/v1 | https://api.stdcmpt.com/v1 |
| Claude Code base URL | http://localhost:4000 | https://api.stdcmpt.com |
| API key | Your LiteLLM key | Your Standard Compute key |
| Model | The model ID you use today | standardcompute, or any exact model ID |
# OpenAI-compatible clients (OpenCode, Cline, Kilo Code, SDKs)
base_url = "https://api.stdcmpt.com/v1"
api_key = "<your-standard-compute-key>"
model = "standardcompute" # or pin one, e.g. anthropic/claude-sonnet-5.5# Claude Code (Anthropic Messages API, no /v1)
export ANTHROPIC_BASE_URL="https://api.stdcmpt.com"
export ANTHROPIC_AUTH_TOKEN="<your-standard-compute-key>"# Codex CLI, ~/.codex/config.toml
model = "standardcompute"
model_provider = "standardcompute"
[model_providers.standardcompute]
name = "Standard Compute"
base_url = "https://api.stdcmpt.com/v1"
env_key = "STANDARD_COMPUTE_API_KEY"
wire_api = "responses"Run your agents on a fixed monthly budget
Every plan includes the compute and every model, and the router spends the budget on the most cost-efficient model for each step. Pay-as-you-go top-ups are optional, never automatic.
FAQ
Is LiteLLM free?
The software is open source under the MIT license, and Enterprise features need a license from BerriAI. You still pay every model provider, plus the cost of hosting and running the proxy.
Does LiteLLM route to the cheapest model?
Its cost-based strategy picks the cheapest deployment within one model group, and its auto router sends prompts to tiers that you map to models by hand. Standard Compute's router picks the model per request.
Does LiteLLM work with Claude Code and Codex?
Yes. LiteLLM documents Claude Code against its /v1/messages endpoint and Codex against /v1/responses. Standard Compute serves both natively without a proxy to run.
Can I keep LiteLLM and add Standard Compute?
Yes. Standard Compute is OpenAI-compatible, so you can add it to LiteLLM as one more provider with base URL https://api.stdcmpt.com/v1 and model standardcompute.
What was the LiteLLM PyPI incident?
On March 24, 2026, LiteLLM versions 1.82.7 and 1.82.8 on PyPI shipped a credential stealer (GitHub issue 24512). Self-hosting means pinning versions and checking releases like any other dependency.
Sources: LiteLLM auto routing · LiteLLM with Claude Code · LiteLLM security issue 24512