Skip to content
Agents tracked: 484 Releases (7d): 690 Agent status: 2 with issuesSince last limit reset: Claude 7d · Codex 5d For AI agents: Markdown Updated Nov 18, 2026

Langfuse changelog: what's new each month

Every stable Langfuse release summarised by month: the highlights, new features, improvements, fixes and anything you need to act on. 4 months covered; the current month updates daily.

November 2026 · so far

13 releases: 4.73.0 → 4.82.0

November brought expanded gateway provider support, improved analytics and access controls, and new capabilities for Topics and sessions. The month focused on enterprise features, audit logging, and enhancing the AI Gateway with additional upstream providers.

Highlights

  • AI Gateway now supports Amazon Bedrock, Google AI Studio, Google Vertex AI, OpenRouter, xAI, and Vercel AI Gateway as providers.
  • Topics gained cost tracking, summaries, assignments, rules in the public API, and can now filter traces and show as table columns.
  • Sessions received a chronological list view, custom view persistence, and integration into trace peek views.
  • Audit logging expanded to record read-access events for traces, sessions, scores, and MCP tools.
  • RBAC improvements include new AI_GATEWAY organization role and USER API key scope evaluated against policies.

New

  • Amazon Bedrock, Google AI Studio, Google Vertex AI, OpenRouter, xAI, and Vercel AI Gateway gateway providers.
  • Topics summaries, assignments, and rules available in public API, MCP, and in-app agent.
  • Chronological session list view with custom view persistence.
  • Session tab in trace peek view and session links in Halo menu.
  • AI_GATEWAY organization membership role for gateway team members.
  • SUBSTITUTE_OK no-data mode for monitors.
  • GitHub Copilot CLI and app setup instructions for AI Gateway.
  • Searchable organization, project, and account switcher.

Improved

  • AI Gateway rate limits raised to 10k requests/min and 5 GiB/min for Pro/Team/Enterprise tiers.
  • Gateway analytics can be filtered by user and API key, with gateway members viewing their own usage.
  • Per-organization request and byte rate limits at the resolve endpoint.
  • Codex skill setup simplified to an exported key and config.toml file.
  • Skill resource loads now traced; skill size and file count limits made configurable.
  • Sessions customization dialog improved; session filtering in Topics setup flow added.
  • Connections & Models tab reorganized with Models tab hidden when no connections exist.
  • Sign-up email codes now have 10-minute expiration with clearer messaging.

Fixed

  • AI Gateway no longer shows shut-down OpenAI and retired Anthropic models.
  • SSRF attacks prevented in GitHub Enterprise SSO configuration.
  • Resolved prompt size capped during dependency resolution to prevent issues.
  • Gateway provider models table border now fits content properly.
  • Web callouts, blob exports, and storage secrets now require validation when endpoints change.
  • Bulk trace deletion audit events now properly aggregated.

October 2026

28 releases: 4.49.0 → 4.72.0

October brought major expansions to Langfuse's AI Gateway capabilities, Topics feature enhancements, and significant improvements to skills management and evaluation workflows. The month focused on making the platform more powerful for enterprise users while improving the experience for managing AI integrations.

Action needed
  • Self-hosted AI Gateway is now gated behind LANGFUSE_AI_GATEWAY_DEPLOYED flag.
  • Legacy API ingestion limited to scores only for new organizations.
  • Renamed Langfuse Assistant to Halo.

Highlights

  • AI Gateway now supports guided setup, per-connection model controls, and detailed analytics for tracking gateway usage and performance.
  • Topics feature gained automatic summarization, mobile support, and new visualization options including treemap and grid views.
  • Skills system was stabilized and integrated into the MCP protocol, with improved CLI installation and management workflows.
  • Organization-level LLM connections allow teams to share model configurations across projects.
  • Evaluators received an overhaul with decision model support, seven-day execution health tracking, and batch evaluation capabilities.

New

  • Organization-level ingestion overview and analytics dashboards with feature activation counts.
  • Tool filters for session tables and dedicated skills tab on trace details.
  • External media storage integration with upload pool management and concurrency controls.
  • Topics classification system that automatically assigns summaries and enables trace map exploration.
  • Topics rules that can be written as target trace transcripts for more flexible filtering.
  • Prompt cache hit rate display in tracing interface.
  • Unstable read-only Topics API endpoints for programmatic access.
  • Self-hosted government plan for enterprise deployments using lf_gov_ license keys.

Improved

  • Session views now use trace transcripts and include message comparison and continuity features.
  • Evaluator gallery redesigned with score previews and improved visual hierarchy.
  • AI Gateway request limit raised to 16 MiB with configurable execution limits.
  • Organization API keys page now supports name search and gateway key filtering.
  • Experiments allow direct annotation from comparison grids and persistent resizable charts.
  • Skills browsing and editing UX improved with file previews and local file/ZIP imports.
  • Design system overhaul including CodeSection replacement for CodeView, improved Timeline connectors, and refined UI component styling.
  • Security hardened with password hashing using PBKDF2 in FIPS mode and API key expiration support.

Fixed

  • Full metadata JSONPath values now display correctly in session headers.
  • Metadata on traces, observations and scores now properly typed as objects in the API.
  • Organizations without ownership permission can no longer select Owner role.
  • Escape LIKE wildcards in remaining substring searches to prevent SQL injection.
  • Trace list labels now properly cleaned up in Topics.
  • Webhook and PostHog destination changes now correctly require secrets verification.

September 2026

33 releases: 4.27.0 → 4.48.0

September brought major improvements to experiments and evaluations, a new AI Gateway service in Rust, and enhanced support for multimodal inputs and conversation transcripts. The platform continued strengthening its observability features and web interface throughout the month.

Action needed
  • Removed unstable evaluation endpoints from the API.
  • New organizations now have legacy read endpoints rejected; clients must use current API.
  • Default JWT session max age reduced to 14 days.
  • Organization API key creation now gated on admin-api entitlement.

Highlights

  • Rebuilt experiments interface with streamlined comparison, score-focused results grid, and improved filtering capabilities.
  • Launched new AI Gateway service with support for OpenAI Responses, Anthropic Messages, and Vertex AI Gemini requests.
  • Added conversation transcript generation from observations with timing metrics and token estimates.
  • Introduced multimodal input support for evaluations and expanded evaluator capabilities.
  • Redesigned session timeline with observation actions and improved visualization.

New

  • Trace message preview feature for internal users.
  • Conversation history and current turn per thread endpoints.
  • Skill management system with draft tracking and file content retrieval.
  • Metadata key presence operators and filter improvements in the query interface.
  • User-triggered session revocation and improved authentication controls.
  • ClickHouse Native write strategy for trace batching.
  • Incident-session scenario seeder for testing the session timeline.
  • Batch inference telemetry collection per project with transient failure retry logic.

Improved

  • Experiments table now includes search bar, dataset filtering by name, and unified score filters across levels.
  • Evaluator interface enhanced with judge prompt metadata tracking and dedicated execution columns.
  • Web interface refactoring with persistent trace summary strip, improved timeline rows, and better label display.
  • Filter interface improvements including inline suggestions, metadata operators, and unified saved views.
  • Session timeline redesigned with observation actions and improved UI layout.
  • API key caching behavior improved with proper eviction after deletion.
  • LLM connections now support custom base URLs and extra headers for TypeSafe connections.
  • Score alerts now properly linked to filtered scores, and dashboard widgets avoid nested scrollbars.

Fixed

  • OpenTelemetry prototype-chain clobbering issue on v3 resolved.
  • Blob storage sub-second remainder chunks no longer cause silent key collisions.
  • API key column order and visibility collisions prevented.
  • Confusion matrix lookup improved to use tuple keys and avoid delimiter collisions.
  • Trace page panels now properly fill window width.
  • SelectInput text truncation fixed in the web interface.

August 2026

30 releases: 4.3.1 → 4.26.0

August brought major improvements to the in-app agent with background execution becoming the default, expanded evaluation capabilities, and significant progress on the v4 migration including a concrete v3 sunset date of November 16, 2026.

Action needed
  • LANGFUSE_AI_PROVIDER environment variable is now required instead of defaulting to Bedrock.
  • V3 API will sunset on November 16, 2026.
  • Foreground runtime support for in-app agent has been removed.
  • Trace and session bookmarking removed from the UI.

Highlights

  • In-app agent background execution is now the default, with new conversation switching and activity tracking during runs.
  • Major evaluation system overhaul with new UX, payload-backed filters, refreshed templates, and evaluator version restores.
  • V4 migration UI now shows by default in dual-write mode with new features pitch and compatibility deadline framing.
  • Expanded agent tool controls including silent mode, per-conversation approval, and sandbox lifecycle metrics.
  • New session interface with header redesign, metadata display, and enhanced filtering options.

New

  • In-app agent can now start new conversations or switch between them while another is running.
  • Tool approval can now be set once for an entire conversation rather than per-use.
  • Session annotations now display existing count totals.
  • Search bar can now filter by facet name in the filter sidebar.
  • Evaluation system includes payload-backed rule filters and refreshed templates for all-caps, exact-match, and keyword-overlap matching.
  • New annotation queue creation directly from trace and events tables.
  • Scores table now includes chart view visualization.
  • Compact dense timeline renderer for traces with maps-style viewport control.

Improved

  • In-app agent now tracks tool approval decisions and records approval source in the event stream.
  • Agent tool silencing and concurrent run limits added per user and organization.
  • Traces now accurately report observations past the load cap instead of hiding them.
  • Trace timeline improved with better lane measurement and fold playback transport into header overflow.
  • Filter sidebar now shares observed-metadata suggestions for better discovery.
  • Dashboard and widget forms now include all filter options and value suggestions.
  • Session header redesign with improved visibility state persistence.
  • Evaluator execution now tracks SLO metrics and displays execution traces linked to evaluations.

Fixed

  • Unicode escapes in client-side trace downloads and copies now properly decoded.
  • Legacy trace widgets now route through v2 correctly.
  • Background run notices no longer display when not needed.
  • Step-limit truncation in in-app agent now surfaces as explicit message instead of silent success.
  • OpenAI Responses API request input now renders properly in trace chat view.
  • Monitors no longer pause on transient ClickHouse resource errors.

Summaries are written automatically from the official release notes (full changelog ↗); check the original notes before relying on a detail. Langfuse: pricing, features and alternatives · All changelogs