Originally published at devtoolpicks.com
Vercel's AI Gateway made a strong pitch to indie hackers. Zero markup on tokens, one API for hundreds of models, budgets built in. But it lives inside Vercel's ecosystem, your request logs live in their dashboard, and this August the whole gateway category got shaken when Stripe agreed to buy OpenRouter for over $7 billion. If you're routing real production traffic through a gateway, this is a good moment to ask where that traffic should actually live.
Here's the short version. The strongest reason to leave Vercel AI Gateway isn't price. It's control: running the gateway inside your own network, keeping logs on your own disk, or escaping a platform you're otherwise trying to leave.
Quick Verdict
| Tool | Best for | Price | Rating |
|---|---|---|---|
| LiteLLM | Self-hosting, full control | Free (OSS) | 9/10 |
| Cloudflare AI Gateway | Managed, zero ops | Free core features | 8.5/10 |
| OpenRouter | Widest model catalog | ~5.5% fee on credit top-ups | 8/10 |
| Portkey | Observability + guardrails | Free tier, then $49/mo | 7.5/10 |
| Helicone | Logging-first teams | Free to 10K req/mo, then $79/mo | 7/10 |
LiteLLM
LiteLLM is an open source proxy you run yourself. It speaks the OpenAI API format, routes to more than 100 providers, and gives you retries, fallbacks, load balancing, and per-key spend tracking without a bill. Your logs never leave your server. For a Laravel or Node app already running on a VPS, it's one more Docker container.
The license costs nothing, but the operations don't. You patch it, you monitor it, you babysit its Postgres. That's a real cost in hours even when the invoice reads zero. An enterprise tier with SSO and audit logs exists at custom pricing, which solo builders can ignore.
Who should NOT use it? Anyone who doesn't want to run infrastructure. If a weekend of downtime while you're away would hurt, a managed gateway is the safer call.
Cloudflare AI Gateway
Cloudflare's gateway runs at their edge and its core features cost nothing. Caching, rate limiting, analytics, and automatic fallbacks are free with any Cloudflare account. Log storage is where limits appear. The free Workers plan keeps 100,000 logs total, and the $5/month paid plan raises that to 10 million per gateway. Unified billing, where Cloudflare buys the tokens for you, carries a 5% fee on credit purchases with no markup on inference itself.
It's the closest like-for-like swap for Vercel's product. Managed, fast, and cheap.
Who should NOT use it? Builders who want deep request tracing and prompt management. Cloudflare gives you counters and logs, not an observability suite.
OpenRouter
OpenRouter is the biggest catalog in the category, with 400+ models behind one endpoint, and inference passes through at provider list price. Its revenue comes from a roughly 5.5% fee (minimum $0.80) when you top up credits by card. Bring your own provider keys and there's no BYOK fee until $25,000 of monthly usage, which no indie project needs to worry about.
And now the elephant. Stripe is acquiring OpenRouter for over $7 billion. Nothing about the fees has changed yet, and the team says it's business as usual. But companies don't pay $7 billion to leave a product alone, and packaging tends to shift after deals like this, the way Anthropic reshuffled its Claude subscriptions this June. We covered the switching math in our OpenRouter alternatives post.
Who should NOT use it? Anyone whose compliance story can't absorb a vendor change mid-acquisition, or who only ever calls one provider.
Portkey
Portkey is a gateway wrapped in an operations console. Routing, fallbacks, and load balancing come with logs, traces, prompt versioning, and guardrails. The free Developer tier records 10,000 logs a month, enough for a side project. Production costs $49/month for 100,000 logs, then $9 per extra 100,000. There's also an open source gateway you can self-host with the routing features and a basic dashboard.
You're paying for visibility, not for tokens. That's the right trade when you're debugging why an agent burned $40 overnight, something we've written about in runaway agent bills.
Who should NOT use it? Cost-minimalists. If you just need routing, LiteLLM or Cloudflare does it for free.
Helicone
Helicone approaches from the logging side. It's open source, and the hosted Hobby tier is free for 10,000 requests a month with one seat. Pro runs $79/month plus usage. You get request logs, cost dashboards, and session traces that make multi-step agent debugging genuinely easier.
One caveat. The gateway feature sits in the paid tiers on the hosted plan, so treat the free tier as observability only.
Who should NOT use it? Anyone picking primarily for routing. Helicone is a magnifying glass first and a router second.
How Do You Choose?
Three questions settle it. Do you want to run infrastructure? If yes, LiteLLM, if no, Cloudflare. Do you need many models or one? A wide catalog points to OpenRouter, a single provider barely needs a gateway. And is your real problem cost visibility? Then Portkey or Helicone, because routing alone won't answer where the money went.
The Bottom Line
For most indie hackers leaving Vercel AI Gateway, Cloudflare AI Gateway is the move. Same managed convenience, free core, no ecosystem strings. Self-hosters with a VPS already running should pick LiteLLM and own the whole stack. Stay on OpenRouter if catalog breadth is why you came, and just keep one eye on what Stripe does with it. Model prices themselves are the same everywhere, and if you're choosing which model to route to, our AI model comparison pages track live pricing daily.
Found a better option? Tell me on X @devtoolpicks.
Top comments (0)