Langfuse Changelog

Experiments now plot every score across runs, show summary statistics against a baseline, and support item-by-item comparison of candidates filtered by score thresholds.

Project API keys can now be created with selected permissions so each integration gets only the access it needs.

LLM credentials can now be configured once at the organization level, with per-project overrides when needed.

New SDK majors are built for Langfuse v4, adding experiment links, typed metadata filters, and compressed span export. Described as a small upgrade for most users.

The TypeSafe System One adapter can be pointed at a custom base URL to run decision-model evaluators with SubQ Snap or Strands Decider.

The redesigned session view reads user and assistant messages across traces as one continuous conversation, with a sidebar for navigating messages, traces, and tool usage.

Project and organization API keys can be given an expiration date, supporting regular rotation and temporary integrations.

Evaluators can use OpenAI's gpt-6-luna decision model to score observations with choice, score, and yes/no questions.

A new chart shows your organization's units over time, grouped by project or type, with a CSV export. On Langfuse Cloud it appears on the Billing page, while self-hosted deployments get a new Usage page.

TypeSafe's Jev decision model can now be used as a judge in Langfuse evaluators, answering typed questions about each observation and returning calibrated scores at lower cost and latency than an LLM judge.

Attaching an evaluator to a rule now backfills scores for recent historical observations.

The Langfuse Assistant can build datasets or dashboards, or run code over thousands of observations in a sandbox, with runs continuing in the background.

Score and cost alerts can now be created directly from evaluator pages, which also list every alert linked to that evaluator.

Langfuse LLM-as-a-Judge evaluators gain two new capabilities: building multi-message prompts and evaluating multi-modal inputs.

Self-hosted Langfuse deployments, including air-gapped installations, now include a built-in interactive API reference for exploring and testing the API directly from the deployment.

The trace Timeline now fits any trace on one screen and zooms like a map, with colour indicating observation type and names shown when there is room.

Evaluators and evaluation rules can now be created, versioned, and managed through stable, ID-based public APIs.

Previous evaluator versions can be restored as a draft, reviewed, and saved as a new version.

Organization admins can enable feature previews for all current and future members, with global personal opt-outs and per-user controls.

Evaluators can now be created from a gallery of templates covering established scoring approaches for chatbots, topic detection, exact matches, and coding agents.