LLM observability
INFRO vs Helicone
Helicone is an open-source observability platform for LLM applications — logging, tracing, cost tracking, evaluations, prompt management, and a gateway — and for the teams already running on it, it answers the question of what an AI application is doing very well. It is no longer a product a new team can adopt: Mintlify acquired Helicone in March 2026, it runs in maintenance mode, and new signups are closed. Nothing is being switched off — Helicone says security updates, bug fixes and new model support keep shipping, and Mintlify has said it will work with every customer on migrating elsewhere — so a team already on it is choosing a destination on its own timetable rather than against a deadline.
- You are already on Helicone and it is doing its job. A frozen roadmap is a reason to plan a move, not to make one this week — and migrating an observability layer in a hurry is how you lose the history you were keeping it for.
- You want to self-host open-source software on your own schedule, with no sales conversation. INFRO runs in your own infrastructure only under an enterprise agreement. Helicone's repository is still public and Apache-2.0 licensed, but adopting a maintenance-mode codebase is a decision with a horizon on it — the live answer here is Langfuse, which is MIT licensed, documented for Docker, Kubernetes and Terraform, and has kept its roadmap since joining ClickHouse in January 2026.
- Observability is the whole job. You already have routing and provider access solved, nobody is asking for budgets that refuse a request, and adding a control plane would be scope you do not need. Langfuse is the one to look at.
- Evaluations and prompt management are the priority. That was a deeper product surface for Helicone than for us, we do not ship prompt management at all, and Langfuse carries both and is still building them.
- You want to start at zero cost. INFRO's entry plan is paid, with a trial. Self-hosted Langfuse costs nothing but the infrastructure and the operating, if you are willing to run it.
This section is not a courtesy. If one of these describes you, go and do it — a reader who adopts the wrong layer on our advice is worse for us than one we never had.
Where the two actually diverge.
- Recording a decision is not the same as making it
- An observability layer tells you, accurately, what your applications did and what it cost. A control plane also decides what they are allowed to do next: which models a project may call, what it may spend before a budget refuses the request, which route carries the traffic when a provider degrades. Those are different jobs, and only one of them changes an outcome while it is happening.
- One layer, so the records cannot disagree
- When routing and observability are separate products, the router knows where a request went and the logger knows what it looked like, and reconciling the two is your job — usually at the moment you least want a data-modelling exercise. Here the request that was routed is the request that was traced, priced, attributed and checked against policy. One record, no seam.
- Cost you can audit, not only cost you can see
- Both products show you spend. INFRO estimates each request against the vendor's published rate, or — from the Team plan — against the rates you actually negotiated, and on the Business plan checks the month's estimate against your provider's own cost API and states the variance. The useful artefact for a finance conversation is not a dashboard number — it is a number you can defend when the provider's invoice arrives.
- Media is a first-class case
- Observability tooling in this category is built around tokens. A video render measured in seconds and dollars, submitted as a job and delivered by webhook, does not fit that shape naturally — and for media-heavy products that is where most of the spend is.
- What the acquisitions actually tell you
- 2026 took several of this category's independent products: Langfuse to ClickHouse in January, Helicone to Mintlify in March, Portkey to Palo Alto Networks in May, and in August Stripe agreed to acquire OpenRouter. That is about as clear as evidence gets that the problem is real, and it is also why the honest thing for us to say is not that INFRO will never be bought. It is that your provider accounts stay in your name, the API is OpenAI-compatible, your request and cost history exports, and leaving is the two lines you changed to arrive — which is the only part of this paragraph that does not depend on who owns us.
Side by side.
Our column carries dashes too. A comparison where one side is all checks is one nobody believes, and rightly.
| Capability | Helicone | INFRO |
|---|---|---|
| Request logging and traces | yesthe core product, open source | yes |
| Your own provider accounts | yes | yesconnected per environment |
| Estimated cost per request | yes | yesitemised; at your negotiated rates from Team |
| Reconciliation against the provider's own bill | no | partialOpenAI and Anthropic, Business plan; variance stated |
| Routing, retries and failover | partialgateway available | yes |
| Budgets that refuse a request | partiallimits and alerting | yesorg, project, member, key |
| Prompt management and evaluations | yesa product surface of its own; roadmap frozen | nonot built, not planned for v1 |
| Image, video, and audio | partialtoken-shaped | yes |
| Self-hosting | yesopen source, Apache-2.0 | partialenterprise agreement |
| Open to new customers | nosignups closed, March 2026 | noprivate early access |
Our reading of each product's public documentation as of October 2026. We do not quote competitors' prices — theirs change and a wrong number about someone else's business is worse than none. Think a row is wrong? Tell us and we will fix it.
Work out what the difference is worth
Estimate what your current model mix costs at your providers' list rates and the plan it needs, or have us read your real usage and tell you where we would and would not help.