Helicone

Helicone spent three years becoming the default answer to "how do I see what my LLM app is actually doing?" — a YC W23 product that grew on a genuinely clever integration model: instead of wrapping your model calls in a new SDK, you change one URL. Your OpenAI client points at ai-gateway.helicone.ai instead of api.openai.com, and every request gets logged, costed, and timed. Then in March 2026 the story changed: Helicone was acquired by Mintlify, the documentation platform, and the team announced that the service continues "in maintenance mode" — security updates, new models, and bug fixes keep shipping, but the roadmap as a living product is effectively closed. That single fact dominates any buying decision today, and I will not bury it further down the page.

What it actually does

Helicone is two things bolted together. First, an observability layer: request logging across providers, cost and latency tracking, session and trace views for multi-step agents, custom properties for tagging, user-level analytics, alerts and scheduled reports, and a query language (HQL) for slicing the request log. Second, an AI gateway: one OpenAI-compatible endpoint that fronts 100+ model providers with intelligent routing, automatic fallbacks when a provider degrades, response caching, custom rate limits, prompt management (deploy prompt versions without code changes), and pass-through billing via Helicone credits with a claimed 0% markup on provider prices. You can also skip the gateway entirely and run observability-only with your own provider keys — the "bring your own keys" mode — which is how a lot of existing users are wired up.

The integration story remains the best in the category. One line of code in an OpenAI, Anthropic, LangChain, Gemini, or Vercel AI SDK client and requests start flowing. There is an async SDK for teams that refuse to route production traffic through a third party. The platform positions itself around three production pains — provider outages breaking apps, unpredictable bills, and bugs you cannot reproduce — and the feature set maps to those: fallbacks for outages, cost dashboards and caching for bills, session replay and prompt versioning for debugging. Datasets and fine-tuning partnerships (OpenPipe, Autonomi) exist but read as adjacent rather than core.

Scale claims from the founders' announcement are substantial: over 14.2 trillion tokens processed, 16,000 organisations, 33 million end users tracked, and a strong footprint among YC companies. Those are vendor figures from the acquisition post, but they are consistent with the GitHub footprint: the Helicone/helicone repository sits at roughly 6,200 stars under Apache-2.0, and — an important detail given the maintenance-mode announcement — it was still receiving commits in late October 2026, consistent with the "bug fixes keep shipping" promise.

Pricing context is worth laying out, because Helicone's positioning always leaned on being cheaper than the SDK-based incumbents. Their own comparison table — vendor-authored, so treat the checkmarks accordingly — prices LangSmith at $39/seat, Langfuse at $59/month, and OpenRouter at a 5.5% provider markup, against Helicone's 0% markup and flat subscription. The performance claim is that edge deployment keeps proxy overhead under 50ms; I cannot independently verify that number without production traffic, but in every functional test the proxy route behaved like a direct call.

My testing notes

I checked the product's current surface area rather than reviewing it as a new buyer would have a year ago. First, the marketing site and pricing page load normally and the app signup URL (us.helicone.ai/signup) returned HTTP 200 for me in late October 2026. That matters because several third-party directories now claim new signups are disabled; I could not verify that claim — a returning 200 on the signup route at least contradicts a hard shutdown, though I did not create an account to test onboarding end to end. Treat "signups disabled" as unconfirmed either way. The homepage served in about 0.4 seconds to first byte from my connection; nothing about the front end suggests a product being switched off.

Second, I pulled the repository metadata: last push 20 October 2026, 170 open issues, 6,234 stars. An open-source LLM observability platform that keeps committing after an acqui-hire-style acquisition is the best reliability signal available here. If the hosted product is frozen forever, the Apache-2.0 codebase is still a viable self-host path, and the self-host stack (ClickHouse-backed, documented in the repo) is the hedge I would build my plan around.

Third, the pricing page itself. The Hobby tier is free forever at 10,000 requests per month with 1 GB of storage and one seat; Pro is $79 per month with unlimited seats, alerts, reports, and HQL; Team is $799 per month with five organisations, SOC-2 and HIPAA compliance, and a dedicated Slack channel; Enterprise is custom with SAML SSO, on-prem deployment, and custom MSAs. Usage-based pricing applies above the included 10,000 requests on all paid tiers. In my experience that free tier is enough to instrument a side project or a staging environment completely, which is exactly the motion that made Helicone popular in the first place.

I also noticed an inconsistency worth flagging: the docs' own comparison table prices Helicone at "$20/seat" while the pricing page sells Pro as a flat $79/month plan with unlimited seats. The maintenance-mode transition is the likely explanation — pages are drifting out of sync as the team moves on — but a buyer reading only the docs would budget wrong. Similarly, the pricing page footer currently carries a "Bifrost" brand mark, a naming change I could not trace to any announcement.

Limits

The big one is direction: there is none. Maintenance mode means security updates and model additions, not new features. If you need evaluation tooling that improves over time, or a platform roadmap you can plan quarters around, this is no longer that product. Helicone's own comparison table admits "limited" evaluation, tracing, and user tracking versus dedicated platforms — those gaps will now never close. The migration posture is a promise, not a guarantee: the announcement says the team will work closely with customers on migrating elsewhere, but there is no published end-of-life date and no contractual commitment to hosting terms, which is an uncomfortable limbo for production dependency. Self-hosting insulates you from the hosting question but takes on the operational burden of running the ClickHouse stack yourself. And while the gateway claims 0% markup on credits, you are consolidating provider billing through a company whose engineering team now works on a different product — a concentration risk that did not exist in 2025. Finally, the docs-versus-pricing contradiction above is a reminder that the information surface is starting to rot.

Verdict

What Helicone built is still, functionally, one of the cleanest LLM observability stacks available: one-line integration, real cost controls, a capable gateway, and a codebase you can take home. The problem is entirely about where it sits in its lifecycle. For new deployments I would treat the hosted product as a short-term instrument — great for visibility now, with a dated migration plan attached — and I would weigh Langfuse (also open source), LangSmith, or Arize as the long-term home. Existing users should export their data regularly and keep the self-host path warm. Rated 3.6 on the rubric: the functionality and usability are high marks, but value and reliability have to be discounted for a product in maintenance mode under new ownership. It was a 4.2 tool in 2025. Context is part of the rating, and I would rather give you a number that survives contact with your procurement team than a nostalgic one. If the hosted platform is still running unchanged in 2027, the codebase alone justifies the score; if Mintlify publishes a sunset date, subtract another half point.

Sources

• https://www.helicone.ai/blog/joining-mintlify
• https://mintlify.com/blog/mintlify-acquires-helicone
• https://helicone.ai/pricing
• https://docs.helicone.ai/getting-started/platform-overview
• https://docs.helicone.ai/gateway/overview
• https://github.com/Helicone/helicone
• https://helicone.ai/blog/introducing-ai-gateway
• https://www.helicone.ai/blog/how-to-gateway
• https://panoply.io/tools/helicone
• https://druce.ai/governance/wiki/vendors/helicone