Helicone review: observability first, gateway second

What Helicone is, and what it is not

Helicone (YC W23) started as one-line LLM logging: point your OpenAI client at their proxy and every request, response, latency, and cost lands in a dashboard. It has since grown outward into evaluation, prompt management, and a gateway, but the centre of gravity has not moved. It is an observability product.

That is a real distinction, not a semantic one. A gateway's job is to get the request to the right model and keep it working. An observability platform's job is to tell you afterwards what happened and what it cost. Helicone is stronger at the second, and shopping for it as an OpenRouter alternative sets the wrong expectation.

Pricing

Beyond the free allowance you pay per log and per gigabyte of storage. Their calculator prices 10,000 requests at 0.30 GB of storage at roughly $0.97, which gives a sense of the unit: fractions of a cent per logged request, dominated by storage on verbose payloads. There is a 7-day free trial with no card.

The 10 logs per minute ceiling on Hobby

This is the free tier's real constraint and it is easy to miss. Ten logs a minute is fine for a demo and useless for an agent, which can emit dozens of tool-call turns in that window. Excess is dropped from ingestion, not from your application, so the failure mode is silently incomplete data rather than a broken request. Evaluate on Pro if you are instrumenting anything agentic.

What it does that Portkey and LiteLLM do not

Three columns, three different products that all describe themselves as AI gateways. The one that matters for a purchase decision: cost per end user and session traces are Helicone features that neither of the others gives you without building it. If your board question is "what does each customer cost us to serve", that is the row you are buying.

The gateway, and the repo you should know about

The hosted gateway is real and works. Point an OpenAI client at https://ai-gateway.helicone.ai, use one key, and you get 100+ models with automatic logging, observability, and fallbacks, with no per-provider SDK.

from openai import OpenAI

client = OpenAI(
    base_url="https://ai-gateway.helicone.ai",
    api_key="<HELICONE_API_KEY>",
)

resp = client.chat.completions.create(
    model="claude-sonnet-5",
    messages=[{"role": "user", "content": "hello"}],
)

The main platform repository is a completely different picture: Helicone/helicone is Apache 2.0, 6,085 stars, and was pushed to on 16 August 2026. The company is active and the hosted product is maintained. But if your plan was "adopt Helicone and self-host the gateway later", check the repository before you commit, because the escape hatch is a year-old beta under a licence that changed to GPL-3.0 after the last code change.

When to pick it, and when not to

Pick Helicone if

• You need cost and behaviour per end user and cannot currently produce it.
• You are debugging agents and need session traces, not individual request logs.
• You want prompt versioning and dataset replay alongside the logs.
• Your request volume is modest and your token spend is high, so log-based pricing beats a percentage.
• You want to start free and instrument something small today.

Do not pick Helicone if

• Routing is the requirement. Portkey, LiteLLM, and OpenRouter all do more here.
• You need guardrails or PII redaction. That is Portkey's product, not this one.
• Self-hosting the gateway is the plan. See the repository dates above.
• You make enormous numbers of tiny requests. Model the per-log cost first.
• You are on the free tier with an agent. 10 logs a minute will silently truncate your data.

Questions people ask

Keep reading

Continue through LLM gateways and routing with How to use OpenRouter in Claude Code: the working setup, and what breaks.