LLM observability — two categories, one OpenTelemetry glue
development platforms
traces + evals + prompts + sessions
LangSmith ($39/user/mo)
Langfuse (MIT + 50K free)
Opik (Apache 2.0)
bundled — best for dev loops
gateway / telemetry
traces + metrics + gateway features
Helicone (proxy, MIT)
SigNoz (Apache 2.0)
OpenLLMetry (OSS)
minimalist — pairs with dev platform
scale / dev / lake
RAG drift, zero-copy, long-term
Phoenix (Elastic L2.0)
Arize AX (zero-copy)
TruLens (OSS evals)
RAG + long-term analytics
OpenTelemetry GenAI conventions — the glue
gen_ai.system · gen_ai.request.model · gen_ai.usage.input_tokens · gen_ai.usage.output_tokens
2025 shipped. 2026 production pattern:
gateway (Helicone) + eval (Phoenix) + lake (Iceberg / Arize AX)
sampling at scale — you can't keep 1M traces/day raw
keep 100% errors · keep 100% high-cost · sample 5% success
aggregates always · raw for long-tail debugging
Arize AX claims ~100x cheaper than monolithic APM at scale — zero-copy on your own data lake