Langfuse

Langfuse's POST /scores endpoint now supports batch processing so multiple scores can be submitted in a single request, and a new Google Cloud Storage export target lets blob-storage exports authenticate via service account key or Application Default Credentials. Additional changes include skill import from local files and ZIPs, FIPS-mode Docker base images, and RBAC API-key authorization resolved from SystemRoleAssignments.

Matters for: Teams running Langfuse with high-volume evaluation scoring workflows or exporting trace data to Google Cloud Storage.

Langfuse now supports an external media storage integration, letting trace and observation media be stored outside the default blob storage. This is in addition to related updates like trace bucket classification and session views built from trace transcripts.

Matters for: Teams running multimodal LLM applications that generate large media attachments (images, audio, files) attached to traces.

Langfuse now supports decision-model evaluators for evals as a generally available feature (the experimental flag was removed), and it embeds comments and annotation review directly beside traces and sessions with a unified score-editing state. The AI Gateway also now accepts Anthropic connections in its resolution contract, and the API adds prompt name prefix filtering.

Matters for: Teams building automated LLM evaluation pipelines across multiple model providers, including those routing Anthropic models through the Langfuse AI Gateway.

This release removes the unstable evaluation API endpoints, backfills saved evaluators, adds cost source visibility in trace cost breakdown tooltips, and introduces a preview 'Improved Message Rendering' feature with a normalized I/O parser; it also raises the playground output token limit ceiling to 65535 and fixes several evaluator, auth, and worker bugs. Because the unstable evaluation endpoints are removed, any integrations or scripts calling those API endpoints must be updated to use the current evaluation API.

Matters for: Teams that built automation or CI pipelines against Langfuse's evaluation API endpoints and teams running self-hosted Langfuse with ClickHouse-backed tracing.

Langfuse now lets you manage alerts for evaluators, filter observations with a dedicated filter builder, search the experiments table, and filter datasets by name instead of id; evaluator names are also disambiguated in filters and observation APIs now use a unified 'model' field instead of 'providedModelName'. Score filters for experiments and items are unified across levels, and dashboard widgets default to the measure's natural aggregation with a more reliable v4 read-path toggle.

Matters for: Teams running multiple evaluators and evaluation pipelines across many Langfuse projects who need alerting and filtering to manage evaluator output at scale.