AI Observability in 2026: Two Markets, One Word

Two markets, one word

Search "AI observability" and the results mix two categories with different buyers, different budgets and almost no overlapping features. Sorting them is the first useful thing anyone can do with the term.

• Observing AI. Is the model doing what we think it is? Span-level traces of model calls and tool use, offline and online evals, prompt versioning, token and latency telemetry. Bought by: The team that shipped the LLM feature. 9 tools scored →
• AI doing the observing. Can an agent handle the incident before a human wakes up? Agentic netOps, AI SRE, alert correlation, root-cause narration, autonomous remediation. Bought by: The team that carries the pager. 15 tools scored →

The distinction matters at the point of purchase. A platform team evaluating incident response and an application team debugging a RAG pipeline will both be shown the same vendor list by most comparison content, and only one of them will find what they came for. Choose the market first. The tool second.

The open half was bought inside twelve months

The two most-adopted open tools in this category are no longer independent, and both moved to companies that sell the infrastructure underneath them.

• Langfuse → ClickHouse (16 January 2026). The most widely deployed open-source LLM observability platform, acquired alongside ClickHouse's $400M Series D at a $15B valuation. MIT licence, self-hosting and roadmap committed to publicly at acquisition. Announcement
• Weights & Biases (Weave) → CoreWeave (May 2025). Acquisition completed in a deal valued at about $1.7B, putting the experiment-tracking incumbent and its LLM tracing layer inside a GPU cloud. Announcement

Neither acquisition changed a licence or closed a self-hosting path on the day it happened, and both acquirers said so explicitly. The question worth carrying into a procurement conversation is a slower one: an observability tool owned by a storage company has an obvious reason to make you store more, and that incentive outlives any launch-day commitment. Ask where the data lands and what leaving costs.

What we can measure, and what we cannot

Our placements come from measured inputs. Reach is GitHub stars and disclosed funding; maintenance is the ratio of closed to open issues, pulled per repository from the GitHub API on 2026-07-29. Nine of the fourteen tools in this cluster publish enough to be scored on that basis, and they carry the source and fetch date per figure so the numbers can be re-checked rather than trusted.

Five do not. LangSmith, Braintrust, Galileo, Fiddler AI and HoneyHive have no public repository and no meaningful presence on the enterprise review platforms, which leaves nothing public to measure on the satisfaction axis. They are listed here with what is verifiable, meaning funding, ownership and positioning, and left off the plot, because a number invented to fill a grid cell is read by every visitor as a number somebody measured.

• 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 itself the most useful finding on this page. The open half of this category can be judged on evidence and the closed half can only be judged on its own marketing. Weigh a vendor demo accordingly.

Where a tool sits is not what it is

Helicone appears in this cluster and in our AI gateways cluster, which is deliberate. Architecturally it is a proxy: you change a base URL and it logs every request. Functionally it is bought for cost and latency visibility, which is an observability job. Portkey sits on the other side of the same line. Same proxy architecture, sold as a control plane, and acquired by Palo Alto Networks in May 2026, which makes it a security product now. Category boundaries drawn on request-path position would put those two together and get both wrong.

How to choose

Most-deployed open-source, now ClickHouse-owned: Langfuse. OpenTelemetry-native tracing with ML-monitoring lineage: Arize Phoenix. Already running Comet: Comet Opik. Evaluation as a test suite rather than a dashboard: DeepEval. Tightest integration for a LangChain codebase: LangSmith. Agent runs rather than single completions: AgentOps or HoneyHive. Regulated environment needing an audit trail across classic ML and LLMs: Fiddler AI. The full scored comparison sits in best LLM observability tools 2026.

For the other market, incident response and remediation, start from the AI SRE rankings: fifteen tools scored, with Datadog Bits AI and LogicMonitor Edwin AI currently placed as leaders.

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)

Landscape last verified 2026-07-29. GitHub figures fetched from the API that day; funding and ownership verified against the company or acquirer announcement in each case. Scored data lives in the LLM observability rankings.

What is AI observability?

AI observability is the practice of knowing what an AI system actually did in production, and it covers two distinct markets. The first is observing AI: span-level traces of model calls and tool use, offline and online evaluations, prompt versioning, and token and latency telemetry, answering whether the model is doing what you think it is. The second is AI doing the observing: agentic netOps and AI SRE tools that point models at conventional infrastructure to correlate alerts, narrate root cause and remediate incidents. They share a phrase and almost nothing else, including the buyer.

What is the difference between LLM observability and AI SRE?

LLM observability instruments the model layer: which prompt ran, which tools were called, what came back, what it cost, and whether the output passed an eval. AI SRE instruments infrastructure and uses a model to interpret it: alert correlation, root-cause analysis and automated remediation across services you already run. One is a new problem with new tools, the other is an old problem with an agent pointed at it. Teams frequently need both, and buy them separately.

Which LLM observability tool should we use?

It depends on what you are optimising for. Langfuse if you want the most-deployed open-source option and can live with ClickHouse ownership. Arize Phoenix if you want OpenTelemetry-native tracing with an ML-monitoring lineage behind it. Comet Opik if your team already runs Comet for experiment tracking. DeepEval if evaluation matters more than tracing and you want it shaped like a test suite. LangSmith if you are already writing LangChain, where the integration is tightest. Our scored comparison ranks the nine we can measure and explains the five we cannot.

Is open-source LLM observability still independent?

Largely not, at the top of the category. Langfuse was acquired by ClickHouse on 16 January 2026, announced alongside a $400M Series D at a $15B valuation, with the MIT licence, self-hosting and roadmap publicly committed to at acquisition. Weights & Biases, whose Weave product carries its LLM tracing, was acquired by CoreWeave in a deal valued at about $1.7B, completed in May 2025. Both remain open and self-hostable today; both are now owned by companies that sell the infrastructure underneath them.

Why do some tools have no score on your grid?

Our placements are computed from measured inputs: reach signals such as GitHub stars and funding, and maintenance signals such as the ratio of closed to open issues. LangSmith, Braintrust, Galileo, Fiddler and HoneyHive publish no repository and have no meaningful presence on the enterprise review platforms, so there is nothing public to measure on the satisfaction axis. Rather than invent a number that the grid would present as measured, we list those five with what is verifiable about them (funding, ownership and positioning) and leave them off the plot.