Helicone Review 2026

Key takeaways

• Helicone is an open-source LLM observability platform and AI gateway. Add one line (a base URL change or header) and every request to OpenAI, Anthropic, Gemini and others is logged with cost, latency and metadata.
• The free Hobby tier (10,000 requests a month) is enough to evaluate it. Pro is $79/month, Team is $799/month, and Enterprise is custom.
• Big caveat: Mintlify acquired Helicone in March 2026, and the company says the service stays live in maintenance mode. Security updates, new models and bug fixes continue. Major new features are not promised.
• SOC 2 and HIPAA coverage starts at the $799 Team tier, a 10x jump from Pro. Teams with compliance requirements should budget for that.
• Best for engineering teams that want cheap, fast visibility into LLM spend and behavior. Weaker for teams that want a long-term roadmap of evaluation features.

Helicone is an LLM observability and routing platform built by Justin and Cole Gottdank, who came out of Y Combinator's W23 batch. The pitch was simple: change your OpenAI base URL, point it at Helicone, and get a dashboard of every request, its cost, its latency and who triggered it. No SDK rewrite, no tracing framework to learn. That simplicity is why it spread quickly among startups building on top of foundation model APIs.

The company reports 14.2 trillion tokens processed, 16,000 organizations and more than 33 million end users tracked over three years. Those are self-reported numbers from its own announcement, so take them as directional. On March 3, 2026, Helicone announced it was joining Mintlify, and the team is moving to San Francisco. That changes how you should evaluate it, and I come back to it below.

The audience is developers and technical teams shipping LLM-powered products: AI startups, SaaS teams adding a copilot, and platform engineers who want one place to see model spend. It is lighter than a full enterprise MLOps suite, which is the whole appeal.

Key features

One-line integration and proxy logging. The core experience is a drop-in proxy. You swap the provider base URL and add a Helicone API key header, and requests flow through and get logged. Helicone has also added native integrations over time, including the Vercel AI SDK and n8n, and the older proxy path is now labelled legacy in the repo, with the newer AI Gateway taking over. In practice, setup takes minutes rather than an afternoon, which is a real advantage over tools that need you to instrument code with tracing decorators.

AI Gateway with routing and fallbacks. The gateway sends requests to the best model for your criteria, whether that is cost, speed, accuracy or availability. Automatic fallbacks switch providers when one errors out or rate limits you. There is a separate open-source repo (Helicone/ai-gateway) describing it as a fast, lightweight router. Supported providers shown on the site include OpenAI, Anthropic, Gemini, DeepSeek, Mistral, Groq, Together AI and OpenRouter.

Caching and rate limits. Response caching cuts spend on repeated prompts, which matters for apps with predictable queries. Rate limits let you cap usage per user or per property, so one runaway customer does not drain your budget. Both sit in the gateway layer rather than in your code.

Cost, latency and usage analytics. The dashboard breaks down spend by model, user, prompt and custom property. This is where most teams get their first surprise: a single feature or a single heavy user accounts for most of the bill. User analytics track end users across requests, and Helicone's own materials cite tens of millions of tracked end users.

Sessions and tracing. Multi-step agent runs and chains can be grouped into sessions so you can see the whole workflow rather than a flat list of calls. Helicone's comparison table marks tracing as supported, while it marks Braintrust as limited. I would still say dedicated tracing tools go deeper on nested spans, so test this against your own agent architecture.

Custom properties and HQL. You tag requests with arbitrary metadata (environment, feature, customer tier) via headers, then filter and group by them. HQL, Helicone's query language, lets you write queries against your own logs when the dashboard filters are not enough. Note that a recent security commit in the repo fixed an HQL cross-tenant bypass, which is reassuring in that it was found and closed, and a reminder that the feature is relatively young.

Prompts, playground, scores and datasets. Beyond logging, there is a Playground for replaying and editing requests, prompt management, scoring of responses, datasets built from production traffic, and webhooks. This covers the lightweight evaluation loop: find a bad response, replay it, tweak the prompt, save it to a dataset. It is not as deep as dedicated evaluation platforms like Braintrust.

Alerts and reports. Pro and above include alerts (spend spikes, error rates) and scheduled reports. For a small team without a dedicated platform engineer, an alert that fires before the monthly invoice arrives is worth the price on its own.

Open source and self-hosting. The main repo, Helicone/helicone, is described as an open-source LLM observability platform. It shows roughly 5,500 commits, and the most recent commit I saw was in September 2026, so it is still being maintained. Enterprise plans also list on-prem deployment.

Who is it for

The sweet spot is a seed to Series B AI startup with two to twenty engineers calling OpenAI or Anthropic in production. Think of a team running a customer support copilot that suddenly finds its token bill has tripled and needs to know which customer or prompt is responsible. Helicone answers that on day one, without a platform migration.

It also suits platform engineers at mid-size SaaS companies who want a gateway for provider failover and per-team budgets, and solo developers who want to log a side project for free. The 10,000 request free tier with one seat is generous for prototypes but small for anything with real traffic, and the 10 logs per minute ingestion cap on Hobby will bite quickly.

Who should skip it? Teams that need a deep evaluation workflow with experiment tracking and human review queues will probably prefer Braintrust or LangSmith. Teams that want a broader ML monitoring stack, including traditional model drift, look at Arize. Regulated companies (healthcare, finance, government) can use it, but only from the $799 Team tier. And anyone betting on a steadily growing feature set should pause, given the maintenance mode announcement.

Integrations and ecosystem

Helicone works with the major model providers: OpenAI, Anthropic, Gemini, DeepSeek, Mistral, Groq, Together AI and OpenRouter all appear on the site. Framework integrations include the Vercel AI SDK and n8n, and the one-line setup works with OpenAI-compatible clients in most languages.

There is an API, with rate limits that scale by plan: 10 calls per minute on Pro, 60 on Team and 1,000 on Enterprise. Data export is available on paid plans, and webhooks let you push events elsewhere. Community support runs through GitHub and Discord, with chat and email support on paid tiers and a private Slack channel and dedicated support engineer higher up.

It is a web dashboard product. There is no notable browser extension or mobile app, which is fine for a developer tool.

Pricing and value

Pricing is request-based with plan-level limits on seats, organizations, retention and ingestion.

• Hobby (free): 10,000 requests a month, 1 seat, 1 organization, 7-day retention, 1 GB storage, 10 logs per minute.
• Pro ($79/month): unlimited seats, 1 organization, 1-month retention, 1,000 logs per minute. Includes the first 10K requests, then usage-based charges. Alerts, reports, HQL and the gateway features are included.
• Team ($799/month): unlimited seats, 5 organizations, 3-month retention, 15,000 logs per minute, SOC 2 and HIPAA coverage, and a private Slack channel.
• Enterprise (custom): SAML SSO, custom MSA, configurable or unlimited retention, on-prem deployment, bulk cloud discounts and the highest limits.

A 7-day free trial is available on the paid plans, with no credit card required per the homepage. Usage above the included requests is billed on a calculator basis. The pricing page example showed 0.30 GB of storage at $0.97 and 10K requests at $0.00, so storage costs are modest at small scale.

Two things to verify before buying. First, one third-party article claims Team includes 10 million free requests, but the official pricing table shows 10K free plus usage-based for every paid tier, so trust the official page. Second, a recent commit in the open-source repo removed self-serve Stripe upgrade endpoints and routed upgrade dialogs to the contact page, while the pricing page still shows trial buttons. That suggests upgrades may move to sales-assisted. Confirm at the pricing page before you plan around self-serve billing. Startups under two years old with under $5M in funding, open-source projects and education users may qualify for discounts, according to secondhand reporting, so ask.

Value-wise, Pro at $79 is competitive for a team that needs unlimited seats and alerts. The $799 Team step is steep, and it is mostly paying for compliance and organizations rather than extra analytics.

Strengths and limitations

What it does well

• Fastest time to value. A base URL change and you have cost and latency data. Few competitors match that.
• Cost visibility. Spend by user, model and custom property is the reason most teams adopt it.
• Gateway features in the same product. Caching, rate limits and fallbacks mean one tool covers observability and routing.
• Open source. You can read the code, self-host, and avoid pure vendor lock-in. The active security hardening in the repo is a good sign.
• Free tier. Enough to try it on a real prototype.

Limitations

• Maintenance mode. After the Mintlify deal, the company says it will ship security updates, new models and fixes, but it does not promise major features. The homepage does not mention this beyond a banner linking to the announcement, so many visitors will miss it. For a tool you plan to embed in production for years, that is a real risk.
• Compliance paywall. SOC 2 and HIPAA only from the $799 plan.
• Evaluation depth. Scores, datasets and the Playground cover basics. Braintrust and LangSmith go further on structured evals.
• Tight Hobby limits. 7-day retention and 10 logs per minute make the free plan a trial, not a home.

I did not find independent third-party reviews or complaint data in this research pass, so the limits above come from the vendor's own pages, the repo and the pricing breakdown.

Bottom line

Helicone is a good choice if you are a small to mid-size engineering team that wants LLM cost tracking, logging and a lightweight gateway running today. Setup is trivial, the free tier is real, and the open-source code gives you an exit if things change.

If you need a platform with an ambitious future roadmap, or deep evaluation and enterprise compliance at a lower price, compare Langfuse, LangSmith, Braintrust and Arize before committing. Best use case in one sentence: getting spend, latency and failover visibility on an OpenAI or Anthropic-powered product in under an hour.