Best LLM Observability Tools 2026: Scored Comparison

The ranking

Nine platforms, ordered by overall score. The score combines reach (GitHub stars, disclosed funding) with maintenance (the ratio of closed to open issues), both measured on 2026-07-29. Where a figure is missing it is shown as missing rather than filled in.

1. Langfuse (62). Open-source LLM engineering platform, now part of ClickHouse. 32.1k stars, 90% issue resolution, owned by ClickHouse.
2. Arize Phoenix (55.7). Open-source tracing and evals from an ML-observability incumbent. 10.8k stars, 89% issue resolution.
3. Comet Opik (54.9). Open-source evals and tracing from the experiment-tracking incumbent. 21.0k stars, 87% issue resolution.
4. DeepEval (47.3). Open-source LLM evaluation framework, pytest-shaped. 17.3k stars, 67% issue resolution.
5. W&B Weave (44.7). LLM tracing inside the experiment-tracking platform teams already run. 1.1k stars, 59% issue resolution, owned by CoreWeave.
6. Helicone (42.7). One-line proxy logging for LLM calls. 6.0k stars, 86% issue resolution.
7. OpenLLMetry (42.4). OpenTelemetry-native LLM tracing. 7.3k stars, 71% issue resolution.
8. AgentOps (39.2). Session replay and tracing built for agents, not chat calls. 5.7k stars, 74% issue resolution.
9. LangWatch (35.5). Open-source monitoring, evals and prompt optimisation. 3.4k stars, 64% issue resolution.

Scored comparison

Same data, side by side. A green mark on "issue resolution rate" means 70% or better; on "independent owner" it means the vendor has not been acquired.

Reach and maintenance are different questions

The two axes disagree often enough that reading only one of them will mislead you. DeepEval sits fourth on reach with 17.3k stars and next-to-last on maintenance at 67%. Helicone has a third of DeepEval's stars and resolves 86% of its issues. Weave has the fewest stars on the list at 1.1k and the lowest resolution rate at 59%, while carrying more brand weight than either through Weights & Biases.

Stars record that a project was found once. The closed-to-open ratio records whether anyone is still answering. For a tool you will file bugs against at three in the morning, the second number is worth more than the first.

The five we cannot score

These vendors publish no repository and have no meaningful review-platform presence, so the signals that place every tool above do not exist for them. They are listed with what is verifiable, and left out of the table.

• LangSmith (LangChain). Framework distribution is the moat: most teams meet LangSmith because they were already writing LangChain. $260M total funding, $1.25B valuation (October 2025).
• Braintrust (Braintrust Data). Evals as a CI/CD release gate, on a trace store written for the purpose. $80M Series B at $800M (February 2026). Not the identically named talent marketplace.
• Galileo (Galileo). Continuous scoring and runtime guardrails, sold to enterprises that want the model policed in production rather than only tested before it. $68M total funding.
• Fiddler AI (Fiddler AI). Explainability and governance lineage predating LLMs, covering traditional ML and generative systems in one control plane. $100M total funding.
• HoneyHive (HoneyHive). Agent-shaped evals and observability, generally available since April 2025 alongside $7.4M led by Insight Partners.

That split is the most useful thing on this page. The open half of this category can be judged on evidence; the closed half can only be judged on its own marketing. Weigh a vendor demo accordingly, and ask closed vendors for the numbers their open competitors publish by default.

How to choose

Most-deployed open-source, now ClickHouse-owned: Langfuse. OpenTelemetry-native tracing with an ML-monitoring lineage: Arize Phoenix. Already running Comet: Comet Opik. Evaluation as a test suite: DeepEval. Traces landing in a backend you already operate: OpenLLMetry. Cost and latency visibility with a one-line install: Helicone. Agent runs rather than single completions: AgentOps. A LangChain codebase: LangSmith, accepting that you cannot check its maintenance signal.

For the other half of AI observability, incident response and remediation rather than model tracing, start from the cluster hub.

Sources

• ClickHouse · ClickHouse welcomes Langfuse (Primary source)
• CoreWeave · Completion of the Weights & Biases acquisition (Primary source)
• Arize AI · $70M Series C (Primary source)
• Braintrust · Announcing our Series B (Primary source)
• Fiddler · $30M Series C (Primary source)
• WTF Rankings · LLM Observability (Our scored data)
• WTF Rankings · AI SRE Tools (Our scored data)

Figures fetched 2026-07-29 from the GitHub API; funding and ownership verified against the company or acquirer announcement in each case. Live scored data sits in the LLM observability rankings.

What is the best LLM observability tool in 2026?

On measured data, Langfuse ranks first: 32.1k GitHub stars and a 90% issue-resolution rate, the strongest combination of reach and maintenance in the category. That is not the same as it being right for you. If your codebase is LangChain, LangSmith has the tightest integration. If you already run Comet, Opik costs you no new vendor. If evaluation matters more than tracing, DeepEval is shaped like a test suite instead of a dashboard. Rank first, then check whether the top of the list actually fits the job.

What are the best open-source LLM observability tools?

Langfuse (32.1k stars, MIT, now ClickHouse-owned), Comet Opik (21.0k, Apache-2.0), DeepEval (17.3k, Apache-2.0), Arize Phoenix (10.8k), OpenLLMetry (7.3k, OpenTelemetry-native), Helicone (6.0k, Apache-2.0), AgentOps (5.7k) and LangWatch (3.4k) are all open and self-hostable. The trade to weigh is not licence but ownership: the two most-adopted are now inside infrastructure companies, ClickHouse and CoreWeave respectively.

How do you score these tools?

From measured inputs only. Reach comes from GitHub stars and disclosed funding; maintenance comes from the ratio of closed to open issues, pulled per repository from the GitHub API on 29 July 2026. Every figure carries its source and fetch date in our database, so it can be re-checked rather than trusted. We deliberately do not use G2 or Capterra ratings here: most developer tools in this category have no meaningful presence there, and several of the comparison pages that do carry ratings are written by competing vendors on this same list.

Why are LangSmith and Braintrust not in the comparison table?

Because there is nothing public to measure. Neither publishes a repository, and neither has a meaningful presence on the enterprise review platforms, so the maintenance and community signals that place every other tool simply do not exist for them. Both are significant products, LangSmith on framework distribution and Braintrust on a purpose-built trace store, and both are covered below with what is verifiable: funding, ownership and positioning. What we will not do is invent a score to fill the row, because a number in a scored table reads as a measured one.

Is Helicone an observability tool or a gateway?

Architecturally it is a gateway: a proxy you point your base URL at, which then logs every request. Functionally it is bought for cost and latency visibility, which is an observability job, and that is why it appears both here and in our AI gateways cluster. Portkey sits on the other side of the same line, sold as a control plane and acquired by Palo Alto Networks in May 2026, which makes it a security product now. Judge these tools by the job they are adopted for, not by where they sit in the request path.