OpenRouter became the default way to experiment with AI models: one API key, hundreds of models, a rotating cast of :free models for zero-dollar prototyping. Then two things happened. OpenRouter tightened its free tier (free models are now gated to 50 requests a day unless your account has ever bought $10 of credits), and in August 2026 Stripe announced an agreement to acquire the company. Neither is fatal, but both are good reasons to map your exits before you need one. Here are the eight genuinely free alternatives worth knowing, grouped by the problem they actually solve.
Why people leave OpenRouter
Most migrations trace back to one of three pain points. The 50-requests-a-day cap on :free models kills heavy testing loops. The 5.5% fee charged on every credit top-up (minimum $0.80) annoys anyone scaling paid usage. And some teams simply want to own their gateway — no third-party cloud in the request path for data-residency or compliance reasons. Different pains point at different products, so the picks below are grouped accordingly.
Self-hosted gateways (zero markup, full control)
1. LiteLLM — the free software gateway
LiteLLM is the open-source (MIT) answer to OpenRouter's managed gateway. You run it on your own server — even a $5 VPS — point it at provider keys you already own, and you pay providers directly with no markup and no per-request fee. It speaks the OpenAI API format, so most clients only need a base_url change, and it supports 100+ models plus spend controls, per-key budgets, and fallbacks.
Best for: developers and teams who want an OpenRouter-style unified API without sending traffic through someone else's cloud, and without paying gateway fees.
2. new-api — the self-hosted reseller panel
If your need is closer to "run a multi-key, multi-provider relay panel" than "route one app," the Chinese open-source project new-api (successor to the MIT-licensed one-api) is the most actively developed option, with tens of thousands of GitHub stars. One caveat from 2026 security writeups: the codebase has carried a cluster of disclosed CVEs (auth-bypass, SSRF, SQLi), so sandbox it, restrict egress, and patch aggressively. Great for internal dashboards; not something to expose to the open internet without hardening.
Best for: teams that need a reseller-style panel with channel management and are comfortable maintaining their own security posture.
Managed gateways with no markup
3. Cloudflare AI Gateway — free core features on every plan
Cloudflare AI Gateway sits in front of whatever providers you already use and adds analytics, caching, rate limiting, guardrails, and fallback logic. The important part: core features are free on every Cloudflare plan, and when you bring your own keys it charges 0% markup. It is also the strongest free option for compliance-sensitive work (SOC 2 Type II, ISO 27001, PCI, GDPR).
Best for: apps already on Cloudflare, or anyone who wants routing and observability without paying a fee on top of provider list prices.
4. Vercel AI Gateway — the zero-markup default for Vercel users
Vercel's gateway is explicitly priced at true 0% markup, including bring-your-own-key usage, behind the OpenAI-compatible endpoints your SDKs already speak. If your app deploys on Vercel, this is the path of least resistance: same account, same dashboard, no extra billing relationship, no per-request tax.
Best for: Next.js and Vercel-hosted projects that want OpenRouter-style multi-model access at exact provider list price.
5. Requesty — lightweight routing with a real free tier
Requesty is a newer, lighter gateway with 400+ models and a genuinely usable free tier: $6 in credits to start, and a free tier of 200 requests a day with no credit card required. It charges a 5% markup on paid usage, so it is a bridge rather than a destination — but as a bridge it is one of the easiest OpenRouter swaps on this list.
Best for: prototyping and side projects that hit OpenRouter's 50-a-day wall and just need more room to experiment.
Observability-first gateways
6. Portkey — guardrails and governance
Portkey pairs multi-provider routing (160+ models) with the features production teams actually get paged about: guardrails, fallbacks, spend tracking, and compliance controls. Its open-source gateway build means you can also self-host. The managed free tier covers 10K logs a month, which is plenty to evaluate it on a real workload.
Best for: teams that need policy enforcement, audit trails, and cost governance around their LLM traffic.
7. Helicone — analytics without switching providers
Helicone takes a different angle: keep your existing provider integrations and BYOK, and layer on best-in-class observability — request logging, caching, prompt management, and cost analytics. The free tier includes 10,000 requests a month, and the core is open source.
Best for: teams that already have provider keys and want visibility into usage, latency, and cost without migrating their request path.
First-party free tiers (skip the gateway entirely)
Sometimes the cheapest alternative to OpenRouter is no gateway at all. These vendors give developers free, rate-limited API access directly:
8. Google AI Studio (Gemini) and Groq — the two free-tier kings
- Google AI Studio offers the most generous general-purpose free API in the industry: rate-limited free access to the Gemini Flash family with up to ~1,500 requests a day and 1M-token context, no credit card required. Trade-off: free-tier prompts may be used to improve Google products (exemptions apply in the EEA, UK, and Switzerland).
- Groq serves open-weight models (Llama, Qwen, GPT-OSS) at LPU speed with a published free plan — around 1,000 requests and 200,000 tokens a day per model — and its terms state it does not train on your inputs. Ideal when latency matters.
Two more worth bookmarking: Cloudflare Workers AI gives 10,000 Neurons a day on the free plan across text, image, embedding, speech-to-text and text-to-speech models, plus Vectorize for RAG on the same plan. GitHub Models exposes a free model catalogue and inference API to any GitHub account, and Cerebras hands out $5 in free credits on signup for absurdly fast open-model inference.
How to switch in an afternoon
Most of these speak the OpenAI API format, so migration is usually a config change, not a rewrite:
1. Inventory your OpenRouter usage. Note which model slugs you call and whether you rely on :free models, auto-fallback, or provider routing.
2. Pick your class. Own the gateway (LiteLLM), zero-markup managed routing (Cloudflare or Vercel), or skip gateways for a first-party free tier (Gemini, Groq).
3. Swap base_url and key. For example, pointing an OpenAI SDK at Cloudflare AI Gateway is a one-line base URL change with your existing provider keys.
4. Re-create fallbacks. OpenRouter's automatic provider fallback is a feature you must rebuild — set up fallback chains in your chosen gateway before deleting the old key.
5. Test the free models first. If you used OpenRouter's rotating :free lineup, start with Gemini's free tier or Groq's free plan; both are more generous and more stable than the 50-a-day cap you are escaping.
Which one should you choose?
- "I want full control and zero fees" → LiteLLM, self-hosted on your own VPS.
- "I want managed routing at list price" → Cloudflare AI Gateway (already on Cloudflare) or Vercel AI Gateway (already on Vercel).
- "I just need more free requests" → Google AI Studio's Gemini free tier, or Groq's free plan for speed.
- "I need guardrails and governance" → Portkey.
- "I only need analytics on top of my keys" → Helicone.
OpenRouter is still a fine product — but in 2026, model routing is a solved problem with genuinely free answers. Pick the class that matches your pain point, and you can migrate in an afternoon.