Self-hosting an LLM observability stack looks cheap on day one: one Docker Compose file, a few containers, no per-trace bill. The real cost shows up later, the first time the SDK in your app and the server you run stop agreeing on a version. In October 2026, Langfuse changed exactly that contract.
Key points
- Langfuse's new Python SDK v5 and JS/TS SDK v6 need a self-hosted server at v4.59.0 or newer.
- Setup tutorials written before October 2026 now point teams down an outdated path.
- The install takes an afternoon, but you own the upgrade path for years, so plan it first.
Why teams self-host LLM observability
LLM observability means recording each step of an AI request. That covers retrieval, the prompt, the model call, token usage, latency and a quality score. Langfuse is an open-source platform that does this, and it can run on your own infrastructure.
Teams self-host for two common reasons. Their traces contain customer data, or they run local models and want everything inside their own network.
When I add tracing to something like a support chatbot, the first finding is rarely about the model. More often it shows that retrieval was the slow step, or that a prompt edit changed answers in a way nobody had noticed. Agents make this visibility more important, because one user request becomes many model calls. McKinsey's 2025 State of AI survey found 62% of respondents' organizations at least experimenting with AI agents.
What changed in Langfuse v4 and the new SDKs
The changes landed in October 2026:
- Server v4 is generally available for self-hosting. Langfuse says its Cloud moves to v4-only on November 16, 2026. Self-hosted teams choose their own migration date.
- New SDK majors shipped on October 14, 2026. Python SDK v5 and JS/TS SDK v6 both require server v4.59.0 or newer (Langfuse upgrade guide).
- Deprecated interfaces were removed.
api.legacy.*andapi.ingestionare gone from the new clients. - The data model is observation-first. User ID, session ID, metadata and tags now sit on every observation, not only on the trace. An observation is one recorded step.
- Runtime floors moved. Python SDK v5 needs Python 3.11 or newer. The OpenAI integration no longer supports pre-1.0 packages.