← Back to changelog

October 14, 2026

Python SDK v5 and JS/TS SDK v6

Hassieb Pakzad

New SDK majors built for Langfuse v4, with experiments recorded through OpenTelemetry, compressed span export, and a client trimmed to the current Public API.

Python SDK v5 and JS/TS SDK v6 are built for Langfuse v4. They send spans with the v4 ingestion header for real-time processing, compress span batches with gzip, record experiments entirely through OpenTelemetry, and send scores in batches to POST /api/public/scores instead of the legacy ingestion endpoint. The generated API clients only include the current Public API, so autocomplete no longer suggests endpoints that Langfuse v4 deprecated.

• Re-run an experiment with the same name on the same dataset and the results land in the same experiment, with a direct experimentUrl / experiment_url to the comparison view.
• Filter on typed metadata: both SDKs JSON-encode every metadata value, so numbers and booleans keep their type and nested objects become filterable dotted paths such as a.b.
• Send large traces with less bandwidth, since span batches are gzip-compressed by default.

Upgrade

Both majors require Langfuse Cloud or self-hosted Langfuse v4.55.0 or later. If you self-host Langfuse v3, stay on Python SDK v4 and JS/TS SDK v5. Score timestamps set in the SDK need Langfuse v4.59.0 or later; v4.55.0 to v4.58.x record the time they receive the score instead.

Python SDK v5 requires Python 3.11+ and langchain-core>=1 for the LangChain integration. JS/TS SDK v6 requires Node.js 22+ and @langchain/core@^1 for the LangChain integration.

pip install --upgrade "langfuse>=5"

npm install @langfuse/client@6 @langfuse/tracing@6 @langfuse/otel@6

The migration guides list every breaking change with before/after code, including removed Public API methods, the experiments API replacing dataset run helpers, JSON-encoded metadata values, trace input/output set on the root observation, and api.scores.create / api.scores.delete replacing api.legacy.score_v1 / api.legacy.scoreV1.