Choose OpenRouter if you want a catalogue and zero operations: one key reaches 415 models across 102 provider endpoints, at provider list price plus a 5.5% credit purchase fee. Choose LiteLLM if you want no per-token fee, no third-party data processor, and full control: it is an open-source SDK and proxy in front of 100+ providers, and you supply the keys, the infrastructure, and the patching. LiteLLM shipped 15 releases in the 12 days to 19 August 2026 and disclosed critical CVEs in April 2026, which is the operational reality you are buying into.
- Fees: OpenRouter 5.5% on top-ups. LiteLLM $0 per token, plus your own infrastructure bill.
- Keys: OpenRouter can be one key with no provider accounts. LiteLLM always needs your own.
- Latency: both add a hop. LiteLLM's can be inside your VPC; OpenRouter's cannot.
- Privacy: LiteLLM adds no processor. OpenRouter adds one, at zero retention by default.
- Ops: LiteLLM released 15 tags in 12 days and had a critical CVE wave in April 2026.
- Coverage: OpenRouter reaches models you have no account for. That is its real moat.
They are not the same category
This comparison gets written as if the two are competing products. They overlap on one function, the unified API, and differ on everything else.
| OpenRouter | LiteLLM | |
|---|---|---|
| What it is | A hosted service and a marketplace | An open-source SDK and a self-hosted proxy |
| Who holds the provider keys | OpenRouter, or you via BYOK | Always you |
| Who runs the endpoint | OpenRouter | You |
| Catalogue | Its own, 415 models / 102 endpoints | Whatever you have accounts for |
| Billing relationship | One, with OpenRouter | N, one per provider |
| Failure mode | Their outage is your outage | Your outage is your outage |
Billing
| Token cost | Gateway cost | Total | |
|---|---|---|---|
| Direct to Anthropic | $40.00 | $0 | $40.00 |
| OpenRouter | $40.00 | $2.20 (5.5% top-up fee) | $42.20 |
| LiteLLM self-hosted | $40.00 | your infra, amortized | $40.00 + infra |
LiteLLM has no per-token fee and no licence cost for the open-source proxy. What it has is a fixed cost that does not scale down: a container, a Postgres instance for keys and spend, a Redis for caching and rate limits, and someone on call. At $500 a month of inference, that fixed cost dwarfs OpenRouter's $27.50. At $50,000 a month, OpenRouter's $2,750 dwarfs it. The crossover is somewhere in the low thousands per month, and it moves depending on what your engineering time costs.
LiteLLM Enterprise is quote-only ("Pricing is based on usage. Contact us for a quote"), unlocked by a LITELLM_LICENSE environment variable. SSO is free up to 5 users; past that it needs the licence.
Operational burden, with numbers
This is the row that decides most real deployments, and it is usually written as a shrug. Here is what the last few months actually looked like on the LiteLLM repository.
| Value | |
|---|---|
| Stars | 56,762 |
| Forks | 10,723 |
| Open issues | 4,996 |
| Release tags, 8 to 19 August 2026 | 15 |
| Latest stable tag | v1.97.0, 16 August 2026 |
Fifteen tags in twelve days is a healthy project moving fast. It is also an upgrade treadmill: you cannot sit on a version for a year, because security fixes ride the same train as features.
OpenRouter's equivalent burden is zero, and its equivalent risk is that you have no control over it.
Latency, privacy, and coverage
Latency
Both add a hop; the difference is where the hop is. A LiteLLM proxy in the same region or VPC as your application adds single-digit milliseconds plus whatever your routing logic costs. OpenRouter adds an internet round trip to their infrastructure before the provider round trip even starts, and neither vendor publishes an overhead figure. For streaming chat this is imperceptible. For an agent doing dozens of short tool-call turns per task, it compounds, and it is worth measuring on your own traffic rather than taking anyone's word for it.
Privacy
| OpenRouter | LiteLLM | |
|---|---|---|
| Extra data processor | Yes | No |
| Prompt logging by default | No (zero logging unless opted in) | Your configuration |
| Opt-in logging incentive | 1% usage discount | n/a |
| Per-provider retention filter | Yes, data_collection: "deny" | You choose the provider |
| Regional pinning | Enterprise only (eu./us. endpoints) | Wherever you deploy |
Coverage and tool-calling fidelity
OpenRouter wins coverage outright. It reaches 415 model entries across 102 provider endpoints, including open-weight models on specialist hosts you would otherwise need separate accounts for, and it will fail over between providers serving the same model. LiteLLM reaches whatever you have credentials for, which for most teams is three or four vendors.
Fidelity is the other way around, and it is subtler than either project admits. Both normalize to the OpenAI chat-completions shape, and normalization is lossy at the edges: prompt-cache control, extended thinking parameters, strict tool schemas, and beta headers are where a request that works against the vendor SDK behaves differently through a gateway. LiteLLM maps every provider's errors onto OpenAI exception types, which is genuinely useful and also means a provider-specific failure arrives wearing a generic name. Test tool calling on your own schemas before you commit either way.
Verdict by scenario
| Scenario | Winner | Why |
|---|---|---|
| Solo dev, exploring models | OpenRouter | No accounts, no deploy, 5.5% is noise at this scale |
| Startup, $500/mo inference | OpenRouter | $27.50 a month is cheaper than any hour of ops |
| Scale-up, $20k/mo, platform team exists | LiteLLM | $1,100/mo of fee buys a lot of container |
| Regulated, prompts cannot leave the VPC | LiteLLM | The only option that adds no processor |
| Product where users pick the model | OpenRouter | Catalogue is the feature |
| Need SSO, audit, org-level budgets | LiteLLM Enterprise or Portkey | Governance is licensed, not free |
| Cannot staff a patch cadence | OpenRouter | Read the April 2026 CVE list again |
| Running coding agents all day | Neither, probably | See below |
Questions people ask
Is LiteLLM cheaper than OpenRouter?
Per token, yes: LiteLLM takes no fee, while OpenRouter adds 5.5% on credit purchases. In total cost, only above a few thousand dollars a month. Below that, the infrastructure and engineering time to run a proxy with Postgres, Redis, and a patch cadence costs more than the fee it saves.
Can I use LiteLLM and OpenRouter together?
Yes, and plenty of teams do. LiteLLM treats OpenRouter as one more provider, so you can self-host the proxy for governance, spend tracking, and virtual keys, and route to OpenRouter only for models you have no direct account for. You pay OpenRouter's 5.5% on that slice of traffic and nothing on the rest.
Does LiteLLM support the same models as OpenRouter?
Not the same set. LiteLLM supports 100+ provider integrations but reaches only the models you hold credentials for. OpenRouter exposed 415 model entries across 102 provider endpoints on 19 August 2026, including open-weight models on specialist hosts, with no account needed for any of them.
Which has lower latency, OpenRouter or LiteLLM?
LiteLLM, if you deploy it near your application, because the extra hop stays inside your own network. OpenRouter adds an internet round trip to its infrastructure before the provider call begins. Neither publishes an overhead number, so measure on your own traffic, especially for agents doing many short turns.
Is LiteLLM safe to self-host?
It is widely deployed and actively maintained, and it has had serious vulnerabilities. April 2026 brought a critical OIDC authentication bypass, a privilege escalation via /config/update that could reach RCE, unsalted password hashes, and a SQL injection in the key verification path. All are fixed, all required upgrading promptly. Self-hosting is safe if you patch; it is not safe if you pin a version and forget it.
Sources
Every figure above was read from these pages on August 2026. Vendors reprice without notice; if you find a stale number, tell us.
- LiteLLM documentation SDK vs proxy, 100+ providers, error mapping
- LiteLLM enterprise quote-only pricing, LITELLM_LICENSE, SSO free to 5 users
- LiteLLM security update, April 2026 CVE-2026-35029, CVE-2026-35030, pass-the-hash, fixed in v1.83.0
- CVE-2026-42208 advisory affected v1.81.16-v1.83.6, fixed v1.83.7
- BerriAI/litellm on GitHub stars, issues, release tags read 19 Aug 2026
- OpenRouter FAQ fee structure, zero logging default
- OpenRouter data policies training toggle, regional endpoints