Originally published at devtoolpicks.com
On August 19, 2026, Stripe announced it agreed to acquire OpenRouter, the gateway that routes requests to 400+ models across more than 80 providers. Bloomberg put the price above $7 billion. The announcement talked about economic infrastructure for AI. It said nothing about your API keys, your integration, or the 5.5% fee you pay when you top up credits with a card.
Maybe nothing changes. Stripe has a decent acquisition record, and it has been building AI infrastructure all year (its agent platform already deploys and pays for things on its own). But if OpenRouter sits in your production request path, "maybe" is carrying a lot of weight. I'd rather have an exit mapped before I need one.
The good news? Model routing is a solved problem in 2026. Two of the alternatives below charge no markup at all, and one of them is free software you can run on a $5 VPS. Every price here was verified against the vendor's own pages this week.
Quick Verdict
| Tool | Best For | Price | Rating |
|---|---|---|---|
| LiteLLM | Self-hosters who want full control | Free, MIT-licensed | 4.5/5 |
| Vercel AI Gateway | Managed routing with zero markup | Provider list price, no fee | 4.5/5 |
| Cloudflare AI Gateway | Apps already on Cloudflare | Free core features | 4/5 |
| Portkey | Teams that need guardrails and governance | $49/month Production plan | 4/5 |
| Helicone | Observability-first teams | Free tier, then $79/month | 3.5/5 |
Why Is Everyone Suddenly Looking for OpenRouter Alternatives?
OpenRouter earned its position honestly. One API key, no markup on inference, and access to basically every model that matters. The fee model was clear too. You pay 5.5% (minimum $0.80) when buying credits with a card, and the model prices themselves match what providers charge directly.
So the product isn't the problem. The problem is that a payments company just bought the routing layer, and the official announcement makes zero commitments to existing customers. No promise the API stays stable. No promise the fee stays at 5.5%. No promise the neutral routing stays neutral once the owner has its own commercial relationships with model providers.
That last one matters more than people admit. OpenRouter's whole pitch was neutrality. Its CEO says that philosophy continues under Stripe. It might. But you don't architect around "might."
And there's a quieter reason to shop around. Gateways have caught up. The features that made OpenRouter special in 2024, unified APIs, fallbacks, and spend tracking, are now table stakes offered by companies whose main business isn't the gateway itself.
LiteLLM
LiteLLM is the escape hatch that can't be acquired out from under you. It's an open source proxy, MIT-licensed, that gives you one OpenAI-compatible endpoint in front of 140+ providers. You run it yourself, usually as a small container with a Postgres database next to it.
The free tier is genuinely complete. Virtual keys, per-key budgets, team management, load balancing, RPM and TPM limits, guardrails, and logging integrations for Langfuse, LangSmith, and OTEL. No credit card, no request caps, no per-token fee. Your only costs are the server it runs on and the model providers you call directly with your own keys.
There's an Enterprise tier with custom pricing that adds SSO, RBAC, audit logging, and dedicated support. Solo builders won't need it.
The real cost is operational. You now own a proxy. It needs deploying, updating, and monitoring, and when it falls over at 2 a.m., that's your pager. It's a light burden (the proxy is CPU-bound and runs fine on a modest VPS), but it isn't zero.
Who should NOT use LiteLLM?
Anyone who got into gateways specifically to avoid running infrastructure. If your app is a weekend project and your deploy target is Vercel, adding a self-hosted proxy with a database is a step backward in simplicity. Take one of the managed options below instead.
Vercel AI Gateway
Vercel AI Gateway is the closest thing to a drop-in OpenRouter replacement, and its pricing is blunt. No markup, no platform fee on tokens. You buy credits, you pay the provider's list price, done. Bring your own provider keys and there's no fee on those requests either, though BYOK requires being on the paid tier with purchased credits.
There's a free tier to kick things around, limited to a subset of models with tighter rate limits. Real work means buying credits, and you're responsible for payment processing fees on those purchases, which is worth noting since dodging a purchase fee might be why you're leaving OpenRouter. Enterprise teams can switch to invoiced billing and skip processing fees entirely.
The catch is in the add-ons. Custom reporting costs $0.075 per 1,000 tag writes and $5 per 1,000 report queries. A team-wide provider allowlist runs $0.10 per 1,000 requests, and so does team-wide zero data retention. None of these are forced on you, and per-request versions of the last two are free. But if you flip every switch, "no markup" stops being the whole story.
Despite the name, it's not locked to Vercel hosting. It's a standard API you can call from a Laravel app on a VPS just as easily as from a Next.js app on Vercel.
Who should NOT use Vercel AI Gateway?
Teams that want rich prompt management, guardrails, and evaluation tooling in the same product. The gateway routes and reports. It doesn't try to be your whole LLM ops platform, so if you want that in one place, look at Portkey.
Cloudflare AI Gateway
Cloudflare AI Gateway has the best headline price of the managed options. The core features cost nothing on every plan. Analytics, caching, and rate limiting are free, you bring your own provider keys, and one line of config points your existing SDK at the gateway.
Caching deserves a special mention. If your app sends repeated prompts (autocomplete, classification, anything template-driven), exact-match caching on Cloudflare's edge can cut a real percentage off your model bill. Spend limits and per-key budgets help too, and if you've read what happens when an agent runs unsupervised overnight, you know why budgets aren't optional anymore.
The limits show up at the edges. Log storage caps at 100,000 logs total on the free Workers tier, against 10 million per gateway on Workers Paid ($5/month). Logpush to external storage needs Workers Paid and costs $0.05 per million requests past 10 million a month. Guardrails are billed as Workers AI inference, so heavy guardrail use has a real cost that scales with prompt length. And their unified billing option charges a 5% fee on credit purchases, which is the exact shape of fee you may be trying to leave behind. Skip unified billing, use your own provider keys, and the 5% never applies.
Who should NOT use Cloudflare AI Gateway?
Anyone who isn't otherwise in the Cloudflare ecosystem and doesn't want to be. The product assumes Cloudflare account plumbing, Workers plans for the useful log tiers, and dashboard-first management. If your stack lives elsewhere, Vercel's gateway will feel less like a second platform to learn.
Portkey
Portkey is the option for when routing alone stops being enough. The gateway covers the basics (universal API, fallbacks, load balancing, retries), then stacks governance on top. Guardrails, role-based access control, semantic caching, unlimited prompt templates, and proper logs with traces.
Pricing is the clearest of the paid tools. The Developer plan is free forever with 10,000 recorded logs a month, and Portkey itself says it's for prototyping, not production. The Production plan is $49/month for 100,000 logs, with overage at $9 per additional 100,000. Enterprise is custom and adds VPC deployment, SSO, and compliance certifications.
Note that you're paying per recorded log, not per token. A chatty app with lots of cheap requests burns through the quota faster than an app making fewer, heavier calls. At indie scale that math usually works out fine. At 1 million requests a month you'd pay $49 plus $81 in overage, about $130, which is still less than a 5.5% fee on any serious model spend.
Semantic caching is the sleeper feature. Unlike exact-match caching, it catches prompts that are worded differently but mean the same thing, which suits user-facing chat where nobody phrases anything identically.
Who should NOT use Portkey?
A solo hacker who just wants one API key for many models. You'd be paying $49/month for governance features built for teams. LiteLLM gives you the routing for free, and Vercel gives it to you managed with no subscription.
Helicone
Helicone comes at this from the opposite direction. It started as LLM observability and grew gateway features (caching, rate limits, automatic fallbacks), so you adopt it when understanding your LLM traffic matters more than routing it.
The free Hobby tier includes 10,000 requests a month and 1 GB of storage, enough to instrument a side project properly. Pro is $79/month with unlimited seats, alerts, reports, and a query language for slicing request data, with usage-based pricing past the included allowances. Team jumps to $799/month for multiple organizations and SOC 2 and HIPAA compliance, which is enterprise territory.
For debugging, it's genuinely good. Full request and response logging, cost per user, cost per feature, and session traces that show you exactly which prompt version started eating money. When you're deciding whether to route traffic to a cheaper model, this is the data that tells you if quality held up. Pair it with a real model comparison and you're making routing decisions on evidence instead of vibes.
As a pure OpenRouter replacement, though, it's the weakest fit here. The gateway exists in service of the observability product, and at $79/month for production use, you're paying observability prices for routing you can get free elsewhere.
Who should NOT use Helicone?
Anyone whose main need is routing. If you'd describe your problem as "one key, many models, don't overcharge me," Helicone solves a different problem than the one you have.
How Do You Choose?
Start from what you're actually optimizing for. Cost, control, or insight. They lead to different tools.
If the answer is control, self-host LiteLLM and the acquisition question never applies to you again. If it's cost with zero ops, Vercel's no-markup gateway or Cloudflare's free tier gets you there, and which one depends on where your stack already lives. If you're a team shipping LLM features to real customers, Portkey's $49 buys guardrails and access control you'd otherwise build. And if your model bill is already scary and you don't know why, that's an observability problem, so it's Helicone.
One more angle worth naming. If most of your model usage is you coding with AI rather than your app serving users, a gateway matters less than picking the right subscription. Gateways solve an API problem, not a tooling problem.
Which One Wins?
For most indie hackers, Vercel AI Gateway. Zero markup, no subscription, no server to babysit, and BYOK when you want direct provider relationships. It's the option that most closely matches what OpenRouter promised, without the acquisition asterisk.
If you have a VPS and enjoy owning your stack, LiteLLM wins instead. Free forever means free forever when it's MIT-licensed code on your own box. Nobody can buy that out from under you.
And to be fair, doing nothing is also a choice. OpenRouter works today, and Stripe may keep it excellent. Just know what you'd migrate to, keep your prompts and provider keys portable, and don't let a $7 billion acquisition be the thing that forces an architecture decision on a deadline you didn't pick.
Found a better option? Let me know on Twitter @hafizdev.
Top comments (0)