LiteLLM

LiteLLM's core pitch holds up: one Docker container gives you a virtual-key gateway in front of 100+ LLM providers (OpenAI, Anthropic, Bedrock, Vertex, Azure, and long-tail providers like Cerebras or Nebius) behind a single OpenAI-compatible /chat/completions endpoint, with budgets, rate limits, fallbacks, and Prometheus metrics included in the free tier — not gated behind a paywall (litellm.ai/pricing). It also ships as a plain Python SDK (pip install litellm) for teams that just want provider-agnostic calls without running a proxy at all. The provider table is the widest of any gateway we've checked — 100+ providers each documented individually, including narrower ones like OCI, Heroku, and Databricks (pypi.org/pypi/litellm/json).

Open Source is $0, self-hosted, no license fee, and the vendor states this explicitly rather than gating features behind a trial clock (litellm.ai/pricing). Enterprise — SSO/SCIM, audit logs, multi-region control plane, secret-manager integrations, 24/7 support SLAs — has no published price; it's quoted against your annual gateway request capacity and deployment shape, with procurement via AWS/Azure Marketplace and a 30-day trial license available on request (docs.litellm.ai/docs/enterprise). SSO itself is free for up to 5 users before a license is required (docs.litellm.ai/docs/enterprise). One thing worth knowing before you self-host: the pip install litellm[proxy] extra pulls in litellm-enterprise, a separate package under a proprietary license (PyPI reports LicenseRef-Proprietary), whose enterprise features simply stay inert until you add a LITELLM_LICENSE key (pypi.org/pypi/litellm-enterprise/json, docs.litellm.ai/docs/enterprise). The repo's root LICENSE file is accurate but easy to misread at a glance — MIT covers everything outside the enterprise/ directory, and that directory carries its own commercial license restricting redistribution and modification without a paid seat count (raw.githubusercontent.com/BerriAI/litellm/main/LICENSE, raw.githubusercontent.com/BerriAI/litellm/main/enterprise/LICENSE.md).

Actively pushed to and not archived (github.com/BerriAI/litellm). PR velocity is high: a spot check of merged PRs in the last day alone shows automated provider price-sync bots, new-model cost-map additions, and active Rust-crate feature work landing within hours of being opened (api.github.com/search/issues). 5,247 open issues is a large backlog for a project this size, which is worth knowing if you're expecting fast turnaround on a niche bug rather than a priced support channel.

LiteLLM announced in June 2026 that it's rewriting the gateway's hot path in Rust, claiming 15x throughput (453 to 6,782 req/s), 11x lower memory (359MB to 32MB), and per-request overhead dropping from ~7.5ms to ~0.05ms in the vendor's own harness (docs.litellm.ai/blog/litellm-rust-launch). As of the current beta docs, this is opt-in per model (rust: true in litellm_params) and covers a specific subset: non-streaming /chat/completions on Anthropic and Bedrock (Converse) only, the Anthropic /v1/messages route on Anthropic and Azure AI, Bedrock audio transcription, and OpenAI Responses-API websockets — anything with tool calls, images, response_format, streaming (except /v1/messages), or extended thinking silently falls back to the existing Python path (docs.litellm.ai/docs/proxy/rust_gateway). The standalone all-Rust server (Mode 2, the one that would actually deliver the throughput numbers end-to-end) has no published Docker image yet — you'd build it yourself from source. Treat the benchmark numbers as a ceiling for a narrow, still-growing code path, not as today's typical request latency.

Four CVEs are publicly disclosed against the LiteLLM proxy: an authenticated SSRF/credential-exfiltration bug (CVE-2026-84377, medium, CVSS 6.5, patched in 1.96.2 with backports to 1.88.x), a second SSRF via a user_config parameter that bypassed the existing guard (CVE-2026-59823, medium, patched 1.83.9), a path-traversal arbitrary file write via Skills archive uploads (CVE-2026-59820, medium, patched 1.83.7), and a local file read via OIDC file references requiring admin access (CVE-2026-59819, low, patched 1.83.10) (api.github.com/repos/BerriAI/litellm/security-advisories). All were fixed within the same or an adjacent minor line; none affects a hosted service since LiteLLM only ships self-hosted. The vendor holds a SOC 2 Type II report available through its Trust Center (trust.litellm.ai/controls) and states it runs no telemetry and stores no data on its own servers for self-hosted deployments (docs.litellm.ai/docs/data_security); there's no public bug-bounty program.

LiteLLM ships a new minor version roughly weekly and, as policy effective 29 June 2026, only patches the four most recent stable minor lines — anything older stops receiving fixes entirely, with no long-term-support track (docs.litellm.ai/docs/enterprise). For a self-hosted gateway holding provider credentials, that's a real operational commitment: falling more than a few minor versions behind means the next CVE patch may not backport to your version.

LiteLLM is built by BerriAI, a Y Combinator W23 company founded by Krrish Dholakia and Ishaan Jaffer (ycombinator.com/companies/litellm). Crunchbase lists the team at 1–10 employees (crunchbase.com/organization/litellm). Reported funding is a $1.6M seed round from Y Combinator, Gravity Fund, and Pioneer Fund, alongside a reported $7M ARR figure — both figures come from a single August 2026 Dealroom news item and have not been corroborated against a company filing or press release; treat the ARR number as unverified (app.dealroom.co/news/feed/litellm-raises-1-6m-seed-funding-amid-security-challenges-hits-7m-arr, read from a search result). The vendor's own README and YC page name Netflix, Stripe, Adobe, and NASA among adopters; these are vendor-supplied logos, not independently confirmed customer relationships.