State of Play 2026: LLM Observability

Where the market stands

CoreWeave completed its acquisition of Weights & Biases on May 5, 2025.

Five more deals followed in 2026:

• Langfuse. ClickHouse announced it had acquired Langfuse on January 16, 2026. ClickHouse says the core features stay open source under the MIT license and Langfuse Cloud continues as a standalone service.
• Promptfoo. Promptfoo, which calls itself a security, evaluation and compliance platform, agreed on March 9, 2026 to be acquired by OpenAI, subject to customary closing conditions. Its site now says Promptfoo is part of OpenAI and remains open source.
• Traceloop. Traceloop announced in March 2026 that it was joining ServiceNow and that its technology would become part of ServiceNow's AI Control Tower. Traceloop says its OpenLLMetry instrumentation will remain open source. On May 5, 2026, ServiceNow described the acquisition as completed.
• Galileo. Cisco announced its intent to acquire Galileo on April 9, 2026 and completed the deal on May 22, 2026. Cisco has owned Splunk since March 18, 2024. Splunk's Agent Observability page states that Cisco acquired Galileo.
• Arize. Dynatrace agreed on August 13, 2026 to acquire Arize in a cash and stock transaction valued at $915 million, and completed it on October 1, 2026. Dynatrace says Arize will keep supporting Phoenix, its open-source platform, and AX, its enterprise platform.

The standard underneath

OpenTelemetry defines semantic conventions for generative AI: the names of the spans, metrics and events an instrumented application emits. In June 2026, release 1.42.0 of the main conventions deprecated every gen_ai attribute, metric, event and span there and moved them to a dedicated GenAI repository. The GenAI span conventions carry the status Development. OpenTelemetry's versioning policy says breaking changes may occur while a signal is in development.

Several products accept this telemetry. LangSmith receives spans at an OpenTelemetry (OTLP) endpoint. Arize builds Phoenix on OpenTelemetry with OpenInference instrumentation.

If you want traces you can move between products, instrument with OpenTelemetry. Plan for attribute renames until the conventions are declared stable.

Four kinds of product

This edition groups the products into four kinds.

• Published code you can host yourself. Langfuse is MIT-licensed for its core features. Phoenix is under the Elastic License 2.0, and Arize says it is free to self-host with no feature limitations. Comet's Opik is under the Apache License 2.0, and Comet says the open-source version uses the same codebase as its hosted versions. Each vendor also sells a hosted service.
• Tied to an agent framework. LangChain's LangSmith comes from the LangChain and LangGraph frameworks and also accepts OpenTelemetry traces from other applications.
• Part of an observability platform. Datadog and Splunk each sell a product named Agent Observability alongside their monitoring platforms. Dynatrace now owns Arize.
• Evaluation first. Braintrust's documentation covers testing changes with experiments, building datasets and running evals. Its GitHub Action posts evaluation results as a comment on each pull request.

What each costs where prices are public

• Langfuse Cloud. Hobby is free with 50,000 units a month. Core costs $29, Pro $199 and Enterprise $2,499 a month, each with 100,000 units. Extra units cost $8 per 100,000. A unit is any tracing data point sent to the platform.
• LangSmith. Developer costs $0 per seat with 5,000 base traces a month. Plus costs $39 per seat per month with 10,000 base traces. Further traces are pay-as-you-go. Base traces are kept 14 days and extended traces 180 days, at an extra fee. Enterprise pricing is custom.
• Braintrust. Starter is free with 1 GB of processed data, 10,000 scores and 14-day retention. Pro costs $249 a month with 5 GB, 50,000 scores and 30-day retention. On Pro, extra data costs $3 per GB and extra scores $1.50 per 1,000. Model calls beyond the included credits are billed at token rates.
• Arize AX. Free includes 25,000 spans and 1 GB a month with 15-day retention. Pro costs $50 a month with 50,000 spans, 10 GB and 30-day retention. Enterprise is custom.
• Comet Opik. Free Cloud includes 25,000 spans a month. Pro Cloud costs $19 a month with 100,000 spans and 60-day retention. Extra spans cost $5 per 100,000.
• Datadog Agent Observability. Free covers up to 40,000 LLM spans a month with 15-day trace retention. Pro starts at $160 a month on annual terms with 100,000 LLM spans, and extra spans cost $3.50 per 10,000. Datadog bills only LLM spans, not tool, workflow or retrieval spans.
• Splunk Agent Observability. Splunk's pricing page reads: "Starting at just $100 per month per 1.2 million spans annually, billed annually." It includes 500 million Luna tokens and 5 Luna fine-tuning runs a year. Luna is the set of small language models Splunk uses for evaluations and guardrails.

All figures are list prices as published. The units differ: units, traces, gigabytes, scores and spans. Convert each offer to your own monthly trace volume and retention before comparing. Ask each vendor which model runs the judge evaluations, and who pays for its tokens.

What self-hosting requires

• Langfuse runs a web container and a worker container on PostgreSQL, Redis or Valkey, ClickHouse and S3-compatible blob storage. Some add-on features need a license key.
• LangSmith self-hosting is an add-on to the Enterprise plan. LangChain's installation guide uses Kubernetes, with PostgreSQL, ClickHouse and Redis or Valkey. LangChain recommends blob storage and external storage services for production.
• Phoenix deploys with Docker or Kubernetes.

Budget the people who will run and upgrade these databases, not only the servers.

What a deployment requires

Plan the evaluation work alongside the instrumentation.

1. Instrument one high-risk application end to end, including retrieval and tool calls. Restrict who can read prompts and outputs, which can contain personal data.
2. Build an evaluation set from real production traces, including known failures.
3. Write evaluators for your own definition of a wrong answer. Check the judge's scores against human labels before trusting them.
4. Add the evaluations to the release pipeline so a regressed prompt fails before it ships.
5. Score a sample of production traffic and alert on falling scores. Send hard cases to human review and back into the dataset.
6. Move the most important checks onto the response path as guardrails only once you know their latency.

Questions to settle first

Settle these before choosing a product.

• Must traces stay inside your own network, and who will operate the databases if they do?
• Which observability platform do you already pay for, and does it now include LLM tracing?
• If the vendor was acquired this year, what does the contract say about price, roadmap and the open-source license?
• Does the product accept OpenTelemetry traces, so you can leave without re-instrumenting?
• How many traces a month will you send, what share will you evaluate, and what will the judge model cost?