What it does
Opik is Comet's LLM and agent observability product. It records model calls, tool invocations, and agent steps so developers can inspect executions, evaluate live traces, and track cost, latency, and quality. A failed production trace can become a test case for later changes. Opik is available as an Apache-licensed self-hosted platform as well as a managed cloud service, with separate free and paid Cloud tiers.
Opik is designed to carry a team from a first trace to an ongoing quality loop. Instrument an application with an Opik SDK or a supported integration, then inspect the recorded model calls, tool invocations, and agent steps in the dashboard. If one trace exposes a problem, the docs describe turning it into a test case, running an evaluation suite, and configuring online rules to score incoming traces. The project also includes prompt optimization workflows, which are useful for comparing alternatives but do not replace a developer's judgment about the correct behavior.
The platform supports both self-hosting and Comet Cloud. The open-source route gives teams control of deployment but also makes them responsible for operating the service; Cloud avoids that work and has a free tier with monthly span and retention limits before paid plans. Since Comet lists a separate Model Production Monitoring product alongside Opik, this record is specifically about Opik's LLM/agent tracing and evaluation product, not Comet's general model monitoring offering.
Use the trace view to narrow an issue to a particular request or tool step, then use evaluations to determine whether a change improves the cases that matter. Opik provides visibility and measurement; it does not automatically fix model behavior or application code.
Strengths
- Traces can be converted into evaluation cases, linking incidents observed in production to regression checks.
- The Apache 2.0 platform can be self-hosted, and the Cloud version has a free tier.
- Online evaluations and dashboards include operational measures such as cost, latency, and error rates.
Watch out for
- A self-hosted deployment means the team installs and runs Opik on its own infrastructure rather than using Comet Cloud.
- The free Cloud plan has monthly span and data-retention limits.
Best for
- LLM and agent developers who want to investigate production traces and turn them into repeatable evaluations.
- Teams comparing a self-hosted AI observability platform with a managed cloud option.
Sources
- Open-Source AI Observability Platform | Opik by Comet (Oct 26, 2026)
- Opik Documentation - Open-Source LLM Observability & Optimization | Opik Documentation (Oct 26, 2026)
- Comet | Flexible Pricing Plans, Free for Academics (Oct 26, 2026)
- opik/LICENSE at main · comet-ml/opik · GitHub (Oct 26, 2026)