Stay free if you only need basic features. Upgrade if you need advanced features. Most solo builders can start free.
Typically 20-50ms per request based on Helicone's published benchmarks. For most applications this is negligible since LLM calls themselves take 500ms-30s — meaning the overhead represents less than 5% of total request time. For latency-critical applications making many sequential calls in agent loops, the overhead can compound and become noticeable. Helicone offers an async logging mode that bypasses the proxy entirely for teams where every millisecond counts — you send requests directly to the LLM provider and POST the request/response data to Helicone's logging endpoint afterward, eliminating any proxy overhead while still capturing full observability data.
Helicone has added session tracking that groups related requests together using a Helicone-Session-Id header, but it's primarily designed around individual request observability. You can attach session IDs and parent-child relationships via Helicone-Parent-Id headers to build hierarchical trace trees, but the visualization is less detailed than dedicated tracing platforms. For deep multi-step agent tracing with custom spans, complex tool call hierarchies, and retrieval pipeline visualization, dedicated tracing tools like Langfuse or LangSmith provide richer instrumentation through their SDK-based approaches. Helicone's strength is capturing every LLM call with minimal setup; for full agent workflow tracing, consider pairing Helicone's gateway-level logging with a dedicated tracing SDK.
Helicone focuses on operational observability (cost tracking, caching, rate limiting) with dead-simple proxy integration that takes under 5 minutes. Langfuse provides deeper tracing, evaluation, and prompt management with SDK-based integration that takes longer to set up but captures richer agent context. Helicone is the better choice when cost visibility and operational controls are the priority; Langfuse wins when you need detailed workflow tracing and evaluation pipelines for complex agent applications. The integration models differ fundamentally — Helicone's proxy approach requires no code changes beyond a URL swap, while Langfuse's decorator and callback-based SDK captures arbitrary application steps beyond just LLM calls. Many teams use both together: Helicone at the gateway for cost controls and caching, and Langfuse via SDK for deep tracing and prompt management.
Yes, Helicone is fully open-source under MIT license and can be self-hosted via Docker. The self-hosted version requires running the proxy gateway, a Supabase backend for storage and authentication, and ClickHouse for analytics, plus optional Redis for caching. It's more operationally complex than the cloud version but gives you full data control — important for healthcare, finance, and EU-based teams with data residency requirements. Helicone publishes a docker-compose setup in their GitHub repository (github.com/Helicone/helicone) with deployment documentation. The self-hosted version includes all core features: request logging, cost analytics, caching, rate limiting, and the full dashboard experience. Enterprise customers can also get dedicated support for on-premise deployments.
Helicone supports 20+ providers including OpenAI, Anthropic, Azure OpenAI, Google (Vertex AI and Gemini), AWS Bedrock, Cohere, Mistral, Groq, Together AI, Fireworks AI, OpenRouter, Perplexity, DeepInfra, Replicate, and custom model endpoints. OpenAI and Anthropic have the most seamless one-line integration via dedicated proxy URLs (oai.helicone.ai and anthropic.helicone.ai). Other providers use the universal Helicone-Target-URL header pattern, which works with any HTTP-based LLM API. Cost calculations are pre-configured for major providers and models, with automatic token counting and per-model pricing. Since the proxy simply forwards HTTP requests, adding support for new providers is straightforward — any endpoint accessible via HTTP can be routed through Helicone's gateway.
Start with the free plan — upgrade when you need more.
Get Started Free →Still not sure? Read our full verdict →
Last verified March 2026