Langfuse released version 4.79.0 on 12.11.2026. The release is centered on the project's AI Gateway, which gained three new providers that moved from pre-release to general availability inside the same version, alongside changes to evaluations, topics, audit logging, and a set of fixes and maintenance updates. The full changelog covers v4.78.0 through v4.79.0.

AI Gateway

  • OpenRouter was added as a pre-release gateway provider (@wochinge, #20002), as were xAI (#20004) and Vercel AI Gateway (#20020); all three were then made generally available in #20028.
  • Custom base URLs reached general availability (#20030), and extra headers became available on all deployments (#20058).
  • OpenAI Decisions was routed as its own API format (#20027) and additionally routed through Vercel (#20069); OpenAI Embeddings was likewise split into its own API format (#20033).
  • Connection models can now be filtered by API format, and the toolbar was tidied (#20074).
  • Vertex AI Flex PayGo Gemini calls are now priced through trafficType service tiers (#20062).
  • Fixes and refactors: provider names are kept out of the delete-dialog replay (#20025); provider configs are now called connections rather than credentials (@hassiebp, #20056); compromised emSign certificate paths are distrusted (@maxdeichmann, #20038); cached resolve contexts are evicted when a key owner's org role changes (#20067); the article "the" is now used before an existing ingestion project (#20072); the model catalog cache is versioned and the Decisions test question was named (#20073); the provider request ID is read once per response (#20063); and usage projection is keyed on ApiFormat (#20064).

Other features

  • Monitors gained a SUBSTITUTE_OK no-data mode (@marksalpeter, #19981).
  • Audit logs now record read-access events for MCP tools (@Steffen911, #19759) and stage and relay audit events on self-hosted enterprise installs (#20024).
  • The isAdmin flag was deleted (@jordibosch20, #19582).
  • Default score configs are seeded when a project is created (@annabellscha, #19911).
  • MCP tools were added to manage the default evaluation model (#20032).
  • The worker now exports settings and feature tables to the core data export (@AbdallahAHO, #20011).
  • Topics received a better display of LLM cost (@nimarb, #19889), plus multiple live rules, one-off analyses, a map picker and a list view (@nmtrang29, #19991); a rule's first map is now built through trace dispatch and the nightly refit job (@LysanderKie, #20093).
  • Prompt list endpoint spans are tagged with pagination and filters (@tobi12345, #20075).
  • Trace transcript evaluations can now be scheduled (@marliessophie, #20006).
  • On the native side, JavaScript JSON value semantics are preserved in Rust and OTLP attributes are converted into owned values (@sumerman, #19951 and #19954).

Fixes and improvements

  • Experiments load every evaluation rule scoped to the dataset (#19942).
  • The value is now optional for null filters (@niklassemmler, #20029).
  • Trace dispatch excludes traces whose recorded start is over a year old (#20047).
  • Datasets return the newest open dataset item by id (#20060).
  • A prompt's dependencies are fetched in one query (performance, #20043).
  • Dependencies were bumped to nodemailer 10.4.1 and fast-xml-parser 5.12.0 (#20071).
  • The sessions sidebar row hover size now matches for the last row (#19847).
  • The security dependabot runs twice daily and on code scanning (#20076).
  • AI feature lock icons are placed beside their switches (#20077).
  • The external code-eval endpoint is parsed with the canonical outbound URL parser (#20082).
  • The API reads full metadata for trace contains filters (@nkabardin, #19883).
  • An empty filter param is no longer replaced by the default view (#19994).
  • Dashboards suggest boolean score names in the Score name filter (#20115).
  • The org switcher moved to the sidebar top and both switchers were tidied (#20012).

Chores and maintenance

  • API docs: a filter condition replaces only the query parameter for the same field (@HardMax71, #18845), and the v2 observations filter docs were aligned with server behavior (#20026).
  • uuid was bumped from 1.28.0 to 1.29.0 in /ai-gateway (@dependabot[bot], #19830).
  • Deployment now assumes the AWS deploy role via GitHub OIDC instead of static keys (#20017).
  • The query AST work continued: the v2 observations view is typed as AST expressions (#19908) and the events row selection is compiled through the AST (#19983).
  • Dependency bumps: minimatch 11.0.1 (#20039), next 16.4.5 (#20041), glob 14.0.3 (#20042), react 19.3.1 and next 16.4.6 (#20104), and @google-cloud/storage 8.4.0 with dockerode 5.0.1 (#20031).
  • The worker shares lazy evaluation scheduling for OTEL ingestion (#19950), and a test now covers the dispatcher's events-dispatch queue path (#20048).
  • Pricing entries were added for GPT-6 Cyber, GPT-5.6 Cyber, GPT-5 Search API and GPT Audio (#20015), and the Bedrock matchPatterns for claude-mythos-5 and claude-mythos-5-1 were widened (@langfuse-bot, #20109).
  • The deps-upgrader security maintainer configuration and tracing were updated (#20095).