Bill by unit model, at increasing spans-per-trace
Same trace volume, only the average trace complexity changes. This is where per-span billing quietly outpaces per-trace billing.
| Spans/trace | LangSmith-style | Langfuse-style | Helicone-style |
|---|
The unit model matters more than the sticker price
Every LLM observability platform advertises a free tier and a per-unit rate, but "per unit" means something different at each one. LangSmith-style pricing treats a trace — one full agent run, however many internal steps it took — as the billable unit, so a complex multi-tool agent costs the same to trace as a single LLM call. Langfuse-style pricing bills every span and score individually: each LLM call, retrieval, tool invocation and evaluation inside that same trace is its own unit, so complexity directly multiplies the bill. Helicone-style pricing sits closer to per-request, tracking API calls rather than logical traces.
The practical consequence: if you're building a simple single-call chatbot, all three models land in a similar range. If you're building a RAG pipeline or multi-tool agent that averages 6-10 spans per trace, per-span billing can run several times higher than per-trace billing for identical traffic — not because the platform is more expensive per se, but because it's counting a different, more granular thing. Before picking a tool, measure your own average spans per trace (most SDKs expose this) and run it through the table above rather than comparing headline monthly prices.
How this calculator works
Free LLM tracing/observability cost calculator — model LangSmith (per-trace), Langfuse (per-span+score) and Helicone (per-request) style billing across your real trace volume and see which unit model punishes multi-step agents most.
Frequently asked questions
Why do LangSmith, Langfuse and Helicone cost so differently for the same app?
Because they don't bill the same unit. LangSmith-style pricing counts traces — one billable unit per end-to-end run, no matter how many internal steps it took. Langfuse-style pricing counts spans and scores — every LLM call, tool call, retrieval step and evaluation inside a trace is its own billable unit. Helicone-style pricing counts requests, which is close to trace count for simple apps but can diverge for multi-call agents. A simple one-call chatbot looks similar under all three models; a multi-step agent with 8 internal spans per trace can cost 6-8x more under per-span billing than under per-trace billing for identical usage.
Which observability tool is cheapest for an AI agent?
It depends entirely on how many spans your average trace produces, not on list price alone. If your agent averages 2-3 steps per trace, per-span billing is usually competitive or cheaper than per-trace billing at the same volume. If your agent averages 6+ steps per trace — common for RAG pipelines, multi-tool agents or reasoning chains with retries — per-trace billing usually wins by a wide margin, because the platform doesn't care how many internal calls it took to produce one trace. Run your own average spans-per-trace through this calculator rather than trusting a vendor's headline price.
Are these exact LangSmith / Langfuse / Helicone prices?
No — these are representative, editable presets based on each platform's public billing model (trace-based, span-based, request-based) as of mid-2026, not a live price feed. Observability pricing changes often and varies by contract, retention window and add-ons. Use the presets to understand which unit model fits your traffic shape, then confirm exact numbers on the vendor's current pricing page before budgeting.