OpenRouter vs Direct OpenAI: When the Middleman Earns His Fee
OpenRouter adds roughly 5.5% on credits and a BYOK fee on top of OpenAI's list prices. The fee math at three volumes, what the marketplace buys, and when direct keys win.
For a workload that is mostly OpenAI models, direct is usually cheaper: the same tokens at list price with no marketplace fee on top. OpenRouter earns its roughly 5.5% credit fee when you use what the fee actually buys, which is hundreds of models behind one key, failover across providers, and one consolidated bill. Below is the fee math at three volumes, the honest case for each side, and the point where the middleman stops earning his keep.
What OpenRouter charges
OpenRouter’s pricing is provider passthrough plus platform fees. Credits carry a fee of roughly 5.5% at purchase, and traffic on your own provider keys carries a smaller fee, around 5% of the underlying usage. Those are June 2026 figures; the fee page has current numbers. The tokens themselves bill at provider list rates, so the fee is the whole price of the marketplace.
That transparency deserves credit. Nothing is hidden in a per-token spread, and you can audit the cost of the convenience on one line of your statement.
What the fee buys
The bundle is real: one API key and one bill for 400+ models across OpenAI, Anthropic, Google, Meta, and the open-weight world, routing and failover you do not operate, a unified activity log, and model switching that is a string change. Teams running three providers through one integration get weeks of saved plumbing for a single-digit percentage.
The question this page exists to ask is narrower: what does that bundle do for an app that calls one provider? The wider field, self-hosted routers included, is mapped in OpenRouter alternatives.
The fee math at three volumes
| Monthly inference spend | Credit fee (~5.5%) | Fee per year | BYOK fee (~5%) per month |
|---|---|---|---|
| $500 | ~$28 | ~$330 | ~$25 |
| $2,000 | ~$110 | ~$1,320 | ~$100 |
| $10,000 | ~$550 | ~$6,600 | ~$500 |
At $500 a month, $28 is a rounding error and arguing about it costs more than paying it. At $10,000 a month of OpenAI-only traffic, the arithmetic sharpens: a 5.5 percent fee on a $10,000 month is $6,600 a year for routing an OpenAI-only app may never use. A retry wrapper around a direct key is an afternoon of work, once.
The bring-your-own-key route softens the math without changing its direction: you still pay around 5% for the marketplace to stand between your key and the provider you already chose.
What direct keys give you
Going direct means OpenAI’s list prices with nothing added: GPT-5.5 at $5 input and $30 output per million tokens, GPT-5 at $1.25 and $10, as of June 2026. It also means the platform’s native features with no intermediary in the path:
- The Batch API, 50% off both sides for jobs that can wait up to 24 hours.
- Cached input pricing, 90% off repeated prompt prefixes on GPT-5.
- Projects and scoped keys, with per-key spending limits and native usage dashboards.
- A direct billing relationship, which matters when compliance wants OpenAI’s paper behind every request.
New platform features land on direct keys the day they ship. Marketplaces generally pick them up afterward, sometimes quickly, sometimes not.
The fallback question, honestly
Failover is OpenRouter’s strongest argument, and it is strongest for open-weight models, where the same weights run on many competing hosts and a dead host is genuinely routable. For OpenAI’s proprietary models the story thins: there are at most a couple of upstream sources, so when OpenAI itself has a bad hour, the marketplace mostly has the same bad hour.
What OpenRouter still smooths for single-provider traffic is rate-limit pressure and transient errors, which a good client library with retries and backoff also handles. Pay the fee for resilience you can verify, not resilience implied by a model count. Teams that want routing without the fee self-host it; that trade is the subject of LiteLLM vs OpenRouter.
Key management and billing
This used to be a stronger OpenRouter point than it is in 2026. OpenAI’s platform now does projects, scoped keys per project, and per-key limits natively, which covers the single-provider version of “give each app its own key and budget.” OpenRouter still wins consolidation when there are five providers and finance wants one invoice. With one provider, there is one invoice either way.
When direct is simply cheaper
The decision compresses to one question: are you buying breadth or buying convenience on a single provider?
| Your situation | Better fit |
|---|---|
| 90%+ of traffic on OpenAI models | Direct key |
| Multi-vendor product, models change weekly | OpenRouter |
| Batch jobs and caching drive your cost strategy | Direct key |
| You want failover without operating anything | OpenRouter |
| Compliance needs a direct provider contract | Direct key |
| Sub-$1,000 spend, value your integration time | Either; relax |
An OpenAI-only workload pays a marketplace fee for breadth it never uses. That is the single sentence to test your own traffic against.
When the bill itself is the problem
Both options on this page are meters. Direct keys remove the fee; they do not change the fact that every retry, loop, and long context bills per token. Disclosure: ProxyLLM is our product, and this is the problem it exists for. Codex Hosted runs OpenAI’s official Codex CLI signed in with your own ChatGPT account and serves it as an OpenAI-compatible endpoint, so OpenAI-bound volume bills to the flat subscription: $129 a month plus your plan, no inference markup, and a $3,500 metered month maps to about $229 all-in on a Pro 5x plan, as an estimate rather than a guarantee. The honest caveats: the flat lane is OpenAI-only and returns complete responses rather than streams. The full three-way logic is in ProxyLLM vs OpenRouter.
If your spend is OpenAI-shaped, run the number through the calculator before you settle the marketplace question; the fee debate is about percentages, and the cost-model debate is about multiples.
Frequently asked questions
Is it cheaper to use OpenAI directly or through OpenRouter?
Directly, on raw price. OpenAI tokens cost the same list rate either way, and OpenRouter adds a fee of roughly 5.5% when you buy credits, as of June 2026. The fee buys one key for 400+ models, cross-provider failover, and a unified activity log. A workload that is mostly OpenAI models usually does better on a direct key.
What fees does OpenRouter charge?
Two main ones as of June 2026: roughly 5.5% when you purchase credits, and a smaller fee, around 5% of the underlying usage, when you bring your own provider keys. Inference itself bills at provider list rates with no per-token markup. OpenRouter's fee page has the current numbers.
What do you give up by leaving OpenRouter for direct OpenAI keys?
One key that reaches hundreds of models, failover across providers and hosts, and a single consolidated bill. You gain native platform features in exchange: the Batch API's 50% discount, cached input pricing, projects with scoped keys and spending limits, and a direct billing relationship with OpenAI.
When is OpenRouter worth the fee?
When you actually use the breadth. Multi-model products, fast experimentation across vendors, and apps that fail over between providers get real value from the marketplace, and the roughly 5.5% fee is usually cheaper than building and maintaining that plumbing yourself.