If you run an agent in production, the failure you actually hit is not "the model got dumber." It is this: a run went wrong two hours ago, you have a trace ID and a vague memory, and your observability tool can filter by tags and time but not by what was actually in the trace. You know the agent said something about a refund. You cannot search for the word "refund." So you scroll.
Langfuse's v4.0.0, tagged stable on July 29, 2026, is largely aimed at that exact moment. It had been in preview since March 10, and the headline feature is the boring, load-bearing one: full-text search across trace inputs, outputs, and metadata. Type the word, get the traces. Everything else in the release is downstream of Langfuse finally being fast enough to do that at scale.
What actually shipped in v4#
Three things matter for a founder deciding whether to upgrade or adopt:
- Full-text search, inline in the filter bar. Backed by ClickHouse full-text indexes, you search inputs, outputs, and string metadata with a
matchesoperator that sits alongside your structured filters. This is the difference between debugging a bad run in seconds and reconstructing it from memory. - Monitors and alerts. v4 can watch cost, quality, and latency and notify your team over Slack, webhooks, or GitHub Actions when a metric drifts out of range. That turns "I check the dashboard when I remember to" into a real regression tripwire — the thing that catches a prompt change that quietly doubled your token bill.
- A faster spine: Observations API v2 and Metrics API v2. The whole v4 effort was a scale rewrite; the practical payoff is that the tables load and the APIs answer without the lag that made the v3 UI painful on a busy project.
The release also brought mobile-optimized layouts, tabbed trace views, append-only agent-run events, and same-week support for Sonnet 5 on Bedrock. None of those is the reason to move, but together they signal a tool being actively pushed, not coasting.
The thing that didn't change — and it's the point#
Here is the fact that decides more than any feature: Langfuse is still MIT-licensed and self-hostable, with unlimited events and unlimited users on self-host. ClickHouse acquired Langfuse on January 16, 2026 — part of a $400M Series D that valued ClickHouse at $15B — and the roadmap, the open-source license, and the self-hosting story all stayed put.
That matters because agent traces are some of the most sensitive data you generate: raw prompts, tool arguments, customer inputs, model outputs. The observability decision is really a data-custody decision.
Choosing an observability tool is choosing who holds your prompts and your customers' inputs. Features are the brochure; custody is the contract.
Langfuse vs Braintrust vs Phoenix#
The three tools founders actually weigh sit at different centers of gravity:
- Langfuse is the one-pane pick: tracing, the new full-text debugging, evals, and prompt management in a single open-source app you can run yourself. Cloud pricing is founder-friendly — Hobby is free (50k events/month, 30-day retention, two seats), Core is $29/month with unlimited users, Pro is $199/month, and every paid tier includes unlimited seats, which is unusual in a market that loves per-seat pricing.
- Braintrust is the eval-first pick: if your daily loop is writing evals, scoring prompt variants, and gating deploys on scores, its playground and scoring workflow are excellent. The trade is a proprietary core and a cloud plane that holds your eval data.
- Arize Phoenix is the standards-first pick: open source and OpenTelemetry-native, so your instrumentation follows OTel/OpenInference conventions and isn't locked to one vendor's SDK. Its eval and prompt-management surface is thinner than the other two.
So which one#
If you are a solo founder or a small team and you want tracing plus real debugging plus evals without handing your prompts to a vendor, Langfuse v4 is the default — start free on cloud to prove it out, and keep the MIT self-host option in your back pocket for when data residency becomes a customer requirement. The one honest caveat: self-hosting Langfuse means running ClickHouse, and a serious deployment is a few thousand dollars a month in infra and ops, not a free lunch. For most teams the answer is "cloud until a contract forces self-host," and the nice thing about an MIT tool is that the migration path exists before you need it.
If evaluation is your workflow, look hard at Braintrust before defaulting. If you are already all-in on OpenTelemetry and want your traces vendor-neutral from day one, Phoenix is the cleaner fit. But for the shape of problem most agent builders have — a run went wrong and I need to find it — v4's full-text search is the feature that ends the scrolling, and it ships in a tool you can own. If you're still standing up the eval side of this, our walkthrough on how to evaluate an AI agent's tool use pairs directly with the tracing you just wired up.
