Key findings
- 20 of 28 sources on budgets, keys, and spend tracking are positive. Okta and Zalando run LiteLLM as their company gateway, and a Reddit commenter caps each team key at US$40 a day.
- Putting coding tools behind one gateway is the most common use, described in 19 sources.
- Stability and code quality draw the most concern, raised in 30 sources, followed by performance under heavy traffic in 15. After the October 2026 compromise, several developers on Reddit moved to Bifrost or another proxy.
- The Ruđer Bošković Institute runs its own LiteLLM proxy for its users with budgets and rate limits, and an independent benchmark there measured about 30 ms of proxy overhead. Zalando's own engineering post describes a company-wide LiteLLM proxy serving about 2,000 monthly users on six small pods. Its platform team values extensibility but restarts the proxy every 20,000 requests because of memory leaks. Pfizer's gateway is still described only in a post Pfizer co-wrote with BerriAI.
- Both hosted alternatives changed owners in 2026. Palo Alto Networks closed its purchase of Portkey on May 29, 2026. Stripe agreed to acquire OpenRouter on August 19, 2026, and neither company had announced the close by October 28, 2026.
Of 32 public accounts that describe a gateway choice, 21 chose another tool or their own code and 11 chose or kept LiteLLM. Teams adopt LiteLLM for spend control and for one endpoint that serves every coding tool, and they leave it over stability and performance under heavy traffic. GILD read English-language public accounts published from January 2025 through October 2026.
Why teams put a gateway in front of their models
Teams praise the control LiteLLM gives them over model use. One team put self-hosted Qwen, Gemma, and DeepSeek behind it to track tokens and move inference in-house, with a wider rollout underway. A university chatbot team at PyData Berlin capped student and staff budgets separately. 12 of the 14 individual developers discussing per-key spend tracking praise it. Pfizer's post with BerriAI describes one gateway for every team, application, and agent.
A second group values the single endpoint. Developers point Claude Code, OpenCode, Aider, Roo Code, and Open WebUI at a LiteLLM proxy so every tool reaches every approved model with one key. One calls it a locally hosted OpenRouter that charges nothing for routing.
Where it breaks
Stability leads complaints, including regressions after minor upgrades and features incorrectly gated to Enterprise. Accounts also report MCP and Admin UI bugs. An engineer found unenforced budgets across several deployments. A GitHub issue shows v1.82.3 recording spend without blocking requests.
Traffic growth exposes performance problems. Users report memory growth that forces restarts and throughput falling from about 16 to 9 requests per second on self-hosted GPUs. Provider rate limits caused Kubernetes readiness failures. Pfizer caught a release that halved throughput in CI. BerriAI measured added p99 latency of 257.7 ms on the Python proxy against 0.7 ms on its Rust beta.
What the coding-tool path needs
Coding agents send provider-specific parameters, headers, and caching instructions that a translating gateway must pass through. Engineers report Claude Code failing through LiteLLM with Kimi and with new Anthropic headers, Codex working only with the Responses API disabled. They also report a Bedrock bill of US$37,901.73 when prompt caching underperformed through the chain. Others run Claude Code on open models through a patched proxy with only web search missing.
Why teams leave, or stay
Departing teams cite stability or latency. One moved to OpenRouter after outages, and another chose Portkey's free plan for speed in March 2026, before Palo Alto Networks bought Portkey. Others evaluated Bifrost for latency or memory use. Some replaced LiteLLM with a small custom service to reduce operations work.
Those who stay name control and cost. A developer who serves internal users prefers a LiteLLM instance to paying OpenRouter's fee on credit purchases, 5.5% on Standard and 8% on Business. An engineer on a team that adopted it after internal discussion reports latency its users do not notice and no surprises against the trade-offs the team expected.
What to settle before choosing
Answer these questions before a pilot ends.
- Must prompts and provider keys stay in your network? If not, price OpenRouter's fee against running the gateway.
- Who will own Postgres, Redis, weekly upgrades, and security patches?
- Which coding tools must work through the gateway, and have you tested each?
- What load test must an upgrade pass before production?
Methodology
GILD opened each of the 106 English-language sources on October 12, 2026, October 25, 2026 for the nine added that day, or October 26, 2026 for four new team accounts. All were published between January 2025 and October 2026. They are Hacker News comments and posts, GitHub issues filed by users on BerriAI/litellm, named engineering and personal blogs, an independent benchmark repository, and recorded talks and videos. Hacker News supplies 73 of the 106. Every count in this piece counts those sources, written by people who chose to post in public. GILD cites BerriAI's customer stories, including Pfizer's, as vendor material and leaves them out of the counts. GILD checked product behavior against BerriAI's documentation on October 12, 2026, and searched again for company-wide rollout accounts on October 26, 2026. It checked who owns OpenRouter and Portkey on October 28, 2026.
Products in this piece
Sources
- Public source
- 18
- Official documentation
- 7
- Zalando Engineering: Agentic Engineering at Zalando (Aug 2026)
- Hacker News comment on moving a company pilot of self-hosted models behind LiteLLM (May 17, 2026)
- Hacker News comment on Brokk using LiteLLM as one API over several models (May 8, 2025)
- Alina Dallmann: One API to Rule Them All? LiteLLM in Production, PyData Berlin 2025 (Nov 23, 2025)
- Hacker News comment on deploying LiteLLM Proxy in several places and moving to Bifrost (Nov 19, 2025)
- BerriAI/litellm issue #26672: budget enforcement bypassed in v1.82.3 (Apr 28, 2026)
- BerriAI/litellm issue #19921: performance regression from 1.80.5 to 1.81.x (Jan 28, 2026)
- BerriAI/litellm issue #21046: throughput from about 16 to 9 requests per second through the proxy (Feb 12, 2026)
- BerriAI/litellm issue #15526: cascading readiness failures under provider 429s (Oct 14, 2025)
- Hacker News post: US$38k AWS Bedrock bill through LiteLLM after a prompt caching miss (Apr 28, 2026)
- Hacker News comment on switching from LiteLLM to OpenRouter (Mar 24, 2026)
- Hacker News comment on switching from the LiteLLM proxy to Portkey (Mar 24, 2026)
- Hacker News comment on a LiteLLM instance for internal users over OpenRouter's fee (Apr 9, 2026)
- Hacker News comment on choosing LiteLLM after internal discussion (Sep 12, 2026)
- Hanlin Xiang on dev.to: 5 pitfalls running LiteLLM Proxy in production (Jun 17, 2026)
- BerriAI/litellm issue #25762: the 5-user SSO limit during evaluation (Apr 15, 2026)
- PyPI Stats: litellm downloads (data as of Oct 12, 2026)
- LiteLLM blog: How Pfizer improved LiteLLM gateway performance and resiliency (Sep 11, 2026)
- LiteLLM blog: Benchmarking the LiteLLM Rust AI Gateway (Jul 22, 2026)
- LiteLLM docs: Budgets and rate limits
- LiteLLM docs: Version support
- OpenRouter pricing and plan comparison
- OpenRouter blog: OpenRouter is Joining Stripe (Aug 19, 2026)
- Palo Alto Networks: General availability of Prisma AIRS AI Gateway (Jul 16, 2026)
- Palo Alto Networks Form 10-K for the fiscal year ended Jul 31, 2026: Portkey acquisition completed May 29, 2026