OpenRouter vs Cloudflare AI Gateway vs LiteLLM (2026)
The Short Answer
Stripe’s agreement to acquire OpenRouter, announced the week of August 17, 2026 at a reported price near $7.5 billion, made a lot of teams look at their gateway layer for the first time in a year. Here is the honest comparison.
| OpenRouter | Cloudflare AI Gateway | LiteLLM | |
|---|---|---|---|
| Model | Hosted marketplace + reseller | Hosted proxy in front of your keys | Self-hosted open-source proxy |
| Who holds provider billing | OpenRouter (you buy credits) | You | You |
| Per-token markup | Yes, embedded in credit pricing | None | None |
| Model catalog | Hundreds, dozens of providers | Whatever you have keys for | Whatever you have keys for |
| Failover across providers | ✅ Built in | ⚠️ Limited | ✅ Configurable |
| Caching | ⚠️ Provider-dependent | ✅ Core feature | ✅ Configurable |
| Observability | ✅ Dashboard | ✅ Strong, request-level | ✅ With your own stack |
| Ops burden | None | Near zero | You run it |
| Lock-in risk | Moderate — new corporate owner | Low | Lowest |
| Best for | Breadth, experimentation, one bill | Cost control with existing keys | Control, compliance, scale |
Verified August 24, 2026.
OpenRouter: Breadth and One Bill
What it is: one OpenAI-compatible endpoint brokering to hundreds of models. You buy credits; OpenRouter settles with providers.
Where it wins:
- Model breadth without account sprawl. Trying Claude Opus 5 ($5/$25 per MTok), Grok 4.6 ($2/$6), Gemini 3.7 Flash (introductory $0.75/$3.75 through December 31, 2026) and GLM-5.3 ($1.40/$4.40) in an afternoon means changing a string, not opening four vendor accounts.
- Open-weight model access. Multiple hosts serve the same weights at different prices; the gateway shops for you.
- Procurement simplicity. One vendor, one invoice — often the deciding factor at small companies where finance approval is the real bottleneck.
Where it costs you:
- You do not hold the provider relationship. No enterprise agreements, no direct rate negotiation, no provider-side support escalation.
- A markup exists because a reseller has to make money. At low volume it is irrelevant; at high volume it is a line item worth measuring.
- New ownership. Stripe’s stated rationale is token routing and usage-based billing. That is a coherent thesis, but it means the roadmap now answers to a payments company’s strategy rather than a developer tool’s.
Cloudflare AI Gateway: Control Without Hosting
What it is: a proxy that sits between your app and the providers you already pay. You keep your keys and your billing; it adds caching, rate limiting, retries, analytics and request logging.
Where it wins:
- You keep the provider relationship. Enterprise terms, committed-use discounts and direct support all survive.
- Caching is a first-class feature, and cached responses are the cheapest tokens in existence — free.
- Request-level observability without building it, which is usually the actual reason teams adopt a gateway.
- No per-token markup. You pay providers what you already paid, plus platform costs.
Where it costs you:
- It is not a marketplace. It cannot get you a model you do not have a key for, and it will not shop open-weight hosts on price.
- Cross-provider failover is thinner than a routing-first product’s. Failing from Claude to Gemini mid-incident needs more of your own logic.
- Another vendor in the request path, with its own availability characteristics.
LiteLLM: Maximum Control
What it is: an open-source proxy and SDK giving one OpenAI-compatible interface over 100+ providers. Run it yourself.
Where it wins:
- No third party in the data path. For regulated workloads or strict data-residency requirements, this is frequently the only acceptable answer.
- Zero markup, full transparency. You see and control routing rules, budgets, retries and fallbacks.
- Portability by construction. Your app talks to a schema you control, deployed where you control it.
Where it costs you:
- You own the uptime. Your proxy becomes a hard dependency of every AI feature you ship, and a single-instance deployment is a self-inflicted single point of failure.
- You own the currency. New models land weekly in 2026 — Gemini 3.7 Flash on August 13, Grok 4.6 on August 12, GLM-5.3 on August 14 — and someone has to keep the config honest.
- You build the dashboards. Observability is possible, not included.
How to Choose
Choose OpenRouter if you are optimizing for experimentation speed and procurement simplicity, your volume is low-to-moderate, and one invoice matters more than a few percent of markup.
Choose Cloudflare AI Gateway if you already hold provider keys, want caching and observability without running infrastructure, and want to keep enterprise pricing relationships intact.
Choose LiteLLM if you have compliance requirements, meaningful volume, or a platform team that would rather own a proxy than accept a vendor in the request path.
Choose more than one if you are honest about failure modes. The strongest 2026 pattern is a self-hosted proxy as the primary path with a hosted gateway configured as a fallback — the gateway becomes a hedge against your own infrastructure rather than a dependency.
The Lock-In Test
Whatever you pick, this five-minute test tells you your real exposure:
- Can you change one environment variable and route to a provider directly? If no, you are gateway-locked.
- Does your code use any non-standard request fields? Every one is migration debt.
- Do you log which provider and model actually served each request? Without this, a routing change and a quality regression look identical.
- Do you have a working direct provider key, tested this month? If not, you have no fallback — only a plan for one.
Teams that pass all four can treat any acquisition news as trivia. Teams that fail should fix the test, not panic about the headline.