Top 5 LLM observability and evaluation platforms

Who this is for: teams that need to trace what their agents did, evaluate quality, and show a supervisor the record. Skip it if you only need basic logs for a prototype.

In short: This category consolidated hard in 2026. Three well-known names are gone or folded into bigger companies. For a regulated firm the deciding questions are whether you can host it yourself, whether it speaks the OpenTelemetry standard (so you are not locked in), and whether human review is logged alongside the traces.

Key points

• Langfuse ranks first for regulated teams that must self-host; ClickHouse acquired it in January 2026.
• LangSmith suits LangChain and LangGraph teams, but self-hosting is Enterprise-only and standard traces are kept 14 days.
• Braintrust is built for evaluation rather than monitoring, with no open-source option.
• Humanloop, Helicone and Galileo have been acquired or folded into other companies.
• Ask vendors whether every trace can be exported in OpenTelemetry format if you leave.

How we ranked

The list

1. Langfuse

• Best for: regulated teams that must self-host.
• Why it ranks first: open source with self-hosting, including air-gapped, or cloud regions in the EU, US and Japan. It accepts OpenTelemetry traces and records agent and tool steps. Named financial-services users include Trade Republic (self-hosted) and Ramp. ClickHouse acquired it in January 2026 and said it would stay open source with no licence changes.
• Watch out for: roadmap risk after the acquisition. Its SOC 2 Type II and ISO 27001 cover the cloud product, not your self-hosted copy.
• Pricing: free tier; Core $29, Pro $199 and Enterprise $2,499 a month; overage $8 per 100,000 units.
• What would change its rank: any move away from the open-source core.

2. LangSmith

• Best for: teams building on LangChain or LangGraph.
• Why it ranks here: deep integration with the most widely used production framework, an OpenTelemetry endpoint, and annotation queues for human review. LangChain raised $125 million at a $1.25 billion valuation in October 2025.
• Watch out for: self-hosting is Enterprise-only, standard traces are kept 14 days, and the pull towards the LangChain ecosystem is real.
• Pricing: free developer tier; Plus $39 per seat a month; Enterprise custom.
• What would change its rank: self-hosting below the Enterprise tier.

3. Arize (Phoenix and AX)

• Best for: teams that want an open-source tool to start and an enterprise product to grow into.
• Why it ranks here: Phoenix is open source; AX added native support for the OpenTelemetry conventions for AI systems, including tool calls and results, in August 2026.
• Watch out for: the Pro tier covers 50,000 spans a month, so production use means Enterprise pricing.
• Pricing: AX free tier; Pro $50 a month; Enterprise custom.
• What would change its rank: clearer human-review features.

4. Braintrust

• Best for: teams whose main problem is evaluation, not monitoring.
• Why it ranks here: built around experiments, scoring and human review, with OpenTelemetry ingestion and SOC 2 Type II. It reportedly raised $80 million at $800 million in February 2026, with customers including Notion, Ramp and Dropbox.
• Watch out for: no open-source option, self-hosting only on Enterprise, and no named regulated financial customers found.
• Pricing: free starter; Pro $249 a month; Enterprise custom.
• What would change its rank: a regulated reference customer.

5. Datadog LLM Observability

• Best for: firms already running Datadog for everything else.
• Why it ranks here: agent and tool-call tracing, human annotation and OpenTelemetry ingestion inside the monitoring tool many operations teams already use.
• Watch out for: SaaS only, and pricing was hard to confirm on Datadog's own site after a recent repackaging.
• Pricing: through Datadog; confirm with sales.
• What would change its rank: a self-hosted option, or clearer pricing.

Gone, folded in, or left out

• Humanloop. Acquired by Anthropic; the platform went offline in September 2025.
• Helicone. Acquired by Mintlify in March 2026 and reported to be in maintenance mode.
• Galileo. Acquired by Cisco in 2026 and now sold as Splunk Agent Observability.
• W&B Weave. Now part of CoreWeave, a GPU cloud; its roadmap may follow that business.
• PromptLayer. Strong on prompt management, narrower on full observability.

Analysts are catching up: Gartner published its first Market Guide for AI Evaluation and Observability Platforms in February 2026.

Three questions for any vendor

1. Can we run it entirely inside our own environment, and what features do we lose?
2. If we leave, can we export every trace in OpenTelemetry format?
3. Show us one agent run with its tool calls and the human approval, on one screen.

Common questions

What is the best LLM observability platform for regulated companies?

Langfuse ranks first because it is open source with self-hosting, including air-gapped, accepts OpenTelemetry traces and records agent and tool steps. The watch-out is roadmap risk after its acquisition by ClickHouse.

How much does Langfuse cost?

There is a free tier, then Core at $29, Pro at $199 and Enterprise at $2,499 a month, with overage at $8 per 100,000 units, as checked 17 Oct 2026.

What happened to Humanloop and Helicone?

Humanloop was acquired by Anthropic and the platform went offline in September 2025. Helicone was acquired by Mintlify in March 2026 and is reported to be in maintenance mode.

What should I ask an LLM observability vendor before buying?

Ask whether you can run it entirely inside your own environment and what you lose, whether you can export every trace in OpenTelemetry format if you leave, and to see one agent run with its tool calls and human approval on one screen.

Disclosure: lowob has no commercial relationship with any vendor on this list. Facts checked 17 Oct 2026.