LiteLLM+8 Mixed

GILD Review

GILD's call and the buying facts behind it.

From GILD research

Should we adopt LiteLLM?

Who on your platform team will take LiteLLM's weekly upgrades and security patches?

30-Second TakePick LiteLLM if prompts and provider keys must stay in your own network and someone will patch it.
Best fitPlatform teams that must keep prompts and provider keys in their own network and cap spend per person, team, or app.GILD Review
Consider another workflowIf nobody will run and patch a gateway, evaluate OpenRouter, which Stripe agreed to acquire on August 19, 2026. If you want a governed platform with hosted observability, evaluate Portkey, now Palo Alto Networks' Prisma AIRS AI Gateway.GILD Review

Drawn from 137 sources and 8 operator insights

GILD analysis

LiteLLM Review

Okta and Zalando run LiteLLM as their company gateway, and practitioners praise its spend control and single endpoint for coding tools while reporting regressions after upgrades, latency, and Enterprise-gated features. The free MIT-licensed core is a self-hosted, OpenAI-compatible gateway to 100+ providers, with virtual keys, budgets, and spend tracking. Adopt it with an owner who can operate the proxy through weekly releases and apply security patches promptly.

AI Gateways
AI Costs

The decision

Best fit: Platform teams that must keep prompts and provider keys in their own network and cap spend per person, team, or app.
Consider another workflow: If nobody will run and patch a gateway, evaluate OpenRouter, which Stripe agreed to acquire on August 19, 2026. If you want a governed platform with hosted observability, evaluate Portkey, now Palo Alto Networks' Prisma AIRS AI Gateway.
Adoption condition: Postgres and Redis in place, budgets set to fail closed, the image pinned by digest to a supported minor line with the GHSA-w4c7 fix, such as 1.106.5, and an owner for weekly upgrades.

Basis: LiteLLM documentation, licensing, and published security advisories, Agentgateway's August 2026 benchmark of LiteLLM's Rust mode, and 125 public accounts from people using LiteLLM, including Okta's and Zalando's engineering posts and comments on Reddit. These include an independent proxy-versus-direct benchmark on the Ruđer Bošković Institute's proxy, which measured about 30 ms of added latency and no bottleneck up to 8 concurrent requests.

Buying & deployment factsWhat a team building on it should verify before production traffic grows.

How you buy it

Free, MIT

Self-hosting is free with keys, budgets, and spend tracking. Enterprise bills yearly by capacity.

Source 1, litellm.ai, LiteLLM pricing: LiteLLM pricing Source 2, BerriAI, Enterprise: Enterprise Source 5, GitHub, LICENSE (GitHub): LICENSE (GitHub)

Details

Self-hosting costs US$0 under MIT outside the enterprise directory, including virtual keys, budgets, rate limits, fallbacks, and spend tracking. Enterprise licenses bill annually by gateway request capacity, architecture, and support, with no per-token charge. Buyers purchase directly, through AWS or Azure Marketplace, or through a reseller. A 30-day trial key is free.

Cost drivers

Provider tokens

Each provider bills at its rates. Plan 1 vCPU and 4 GiB per gateway pod.

Source 1, litellm.ai, LiteLLM pricing: LiteLLM pricing Source 6, BerriAI, Production best practices: Production best practices Source 8, BerriAI, Budgets and rate limits: Budgets and rate limits Source 2, BerriAI, Enterprise: Enterprise

Details

Each provider bills tokens at its rates. BerriAI recommends 1 vCPU and 4 GiB per gateway pod, with Postgres for keys and spend and Redis across instances. Enterprise adds SSO beyond 5 users, audit logs, and tag budgets. Round-the-clock support SLAs cost extra.

Where it runs

Your infrastructure

Run it in your cloud. Prompts stay off BerriAI; Enterprise reports successful request counts.

Source 2, BerriAI, Enterprise: Enterprise Source 7, docs.litellm.ai, Production deployment: Production deployment Source 1, litellm.ai, LiteLLM pricing: LiteLLM pricing Source 9, BerriAI, Data privacy and security: Data privacy and security Source 10, docs.litellm.ai, Billable request metering: Billable request metering

Details

LiteLLM runs in your cloud or data center using signed Docker images, a Helm chart, or Terraform for AWS and GCP. Enterprise supports air-gapped deployments. Self-hosted LiteLLM keeps prompts and telemetry off BerriAI servers. Enterprise metering sends a count of successful requests.

What to verify

March 24, 2026

Two PyPI releases carried credential-stealing malware. Only the latest four minor lines get patches.

Source 12, api.github.com, GitHub Advisory Database: reviewed litellm advisories (API): GitHub Advisory Database: reviewed litellm advisories (API)Source 4, docs.litellm.ai, Security update, March 2026: Security update, March 2026 Source 11, docs.litellm.ai, Security incident, v1.107.0 (Oct 21, 2026): Security incident, v1.107.0 (Oct 21, 2026)Source 13, api.github.com, Advisory GHSA-w4c7-34wj-hqm8 (GitHub API): Advisory GHSA-w4c7-34wj-hqm8 (GitHub API)Source 3, BerriAI, Version support: Version support Source 8, BerriAI, Budgets and rate limits: Budgets and rate limits

Details

GitHub lists 44 reviewed litellm advisories as of October 28, 2026, including 7 critical ones. PyPI versions 1.82.7 and 1.82.8 carried credential-stealing malware on March 24, 2026. On October 21, 2026, PyPI 1.107.0 and the Docker tags v1.107.0, latest, main-latest, and main-stable did too. Pin 1.106.5 by digest. Only the four latest weekly minor lines receive patches. Budgets fail open without a database.