OpenRouter and TokenRouter look remarkably alike. Same lineup of supported models. Same OpenAI-compatible API. Even the same model ID convention, right down to anthropic/claude-opus-5. Line up their feature tables side by side and nothing separates them.
What does separate them is two things: which numbers they disclose, and where they put the authority to choose a model. Let’s take them in turn.
OpenRouter — Full fee disclosure, thin org features
OpenRouter’s pricing is unusually plain-spoken. It doesn’t touch per-token rates at all. From the FAQ: “We pass through the pricing of the underlying providers without any markup.”
So where does the money come from? A 5.5% payment fee when you buy credits (minimum $0.80, or 5% for crypto). With BYOK — bringing your own API key — the fee is 5% of what the same model would have cost through OpenRouter, but it’s waived up to a threshold. That allowance is measured in list-price inference cost (dollars), not request count: Pay-as-you-go runs fee-free up to $25,000/month, Enterprise up to $200,000/month (per the official FAQ as of August 21, 2026; 5% applies above those lines). Free models give you 50 requests/day with no credits purchased, or 1,000/day once you’ve bought $10 or more.
Because every number is on the table, you can work out the economics of switching on a calculator. That’s the single best thing about it.
The routing controls lean developer-first too. Provider attempt order is order, exclusions are ignore, restriction is only, and quantizations lets you pin down quantization precision. It’s a mark of honesty that the docs volunteer the downside themselves: “Quantized models may exhibit degraded performance for certain prompts.”
There’s also Fusion, which runs multiple models in parallel — and this is where people misread things. The official docs state explicitly that the judging model does not merge the responses (it doesn't merge them). What it actually does is return a structured JSON comparison covering points of agreement, disagreement, and omissions. With the default three-model setup, that runs four to five times the usual cost. Not a feature for daily use.
The weakness shows up in org features. Two roles — admin and member — and a 10-seat cap per organization. Nothing in the official docs about allocating budgets per department. Fine for a team of five. The moment finance asks for a per-department breakdown, you’re stuck.
One more detail, small but telling. OpenRouter’s list of supported tools still includes “Windsurf.” That product was renamed Devin Desktop on June 2, 2026. Speaking as someone who maintains a catalog, this kind of lag is unremarkable — which is exactly why a supported-tools list should be read for when it was last updated, not for whether something appears on it.
TokenRouter — Heavy on org control, silent on rates
TokenRouter is shaped the opposite way.
Before going further, let me sort out the name. More than one service calls itself “TokenRouter.” The one this article covers, tokenrouter.com, was announced on April 21, 2026 by Silicon Valley AI infrastructure company PaleBlueDot AI as “PBD TokenRouter” (the terms of service and DPA name Delaware entity Artemis Tokenrouter Inc. as the contracting party). Separately there’s tokenrouter.io — an unrelated company with a flat-rate monthly product supporting 13 providers — plus several same-named OSS projects on GitHub. When you read a write-up on this, check which domain it’s about first.
Back to pricing. TokenRouter charges no platform fee. The FAQ says so directly — “TokenRouter currently does not charge platform fees” — and it doesn’t pass Stripe’s payment fees on to users either. On that basis alone it looks cheaper than OpenRouter.
But it doesn’t disclose whether there’s a markup on per-token rates. No markup percentage, and no statement that there is none, anywhere I could find on the official site. Same for the minimum top-up amount and the default RPM/TPM ceilings — undisclosed (raising them means contacting their business team). On top of that, the terms of service define prepaid balances and top-ups as “non-cancellable, non-refundable.” You pay up front, and there are no refunds.
The numbers wobble elsewhere too. The model list page shows “Showing 1 to 24 of 117 entries,” the company blog says “300+ AI models,” and the official feature guide says “50+ multimodal models.” Three figures from three channels at the same company. For reference, OpenRouter’s homepage currently shows “500+ active models on 80+ providers” (as of August 21, 2026; when this article first ran it read “400+ active models on 70+ providers”). The lesson: don’t believe a headline number until you count it somewhere countable.
What you get in exchange is a well-built organizational layer. Usage rolls up to the org account, so nobody fronts costs and files expense reports. Quotas can be allocated per individual, per team, or per department. Analytics break down by model, by member, and by period, and logs go deep enough to trace which request, from whom, consumed how many tokens on which model. It builds head-on for the territory where OpenRouter tops out at 10 seats.
Data retention is billed as “zero retention,” but the scope needs care. Section 7.1(b) of the DPA states plainly that ZDR does not automatically extend to model providers, and that customers should verify each provider’s policy individually. The gateway not retaining your data and everything downstream not retaining it are two different claims.
On routing, note that the resolution varies by source. The launch press release described “Smart Token Routing” as analyzing requests and automatically dispatching them to the model best suited to the task, and also mentioned “Smart Caching” for request deduplication and result reuse, plus 99.95% uptime via multiple paths. The official site and docs, meanwhile, contain no description of those decision criteria or how to configure them. What’s written there stops at “Always-On Routing” and “Dynamic Global Routing” — phrases pointing at infrastructure-layer optimization based on latency and availability.
And governance-oriented policy routing — “inspect the prompt content and block this model if it contains sensitive data,” “pick a provider according to regional regulation” — appears in neither the press release nor the official docs. Control here centers on budgets and quotas, not on what’s inside the prompt. If that’s your reason for considering it, ask them before you commit.
Switching rewrites your base URL — with one bump for Claude Code
On migration cost: what actually changes is the endpoint URL and the key. Both use provider/model for model IDs, written as anthropic/claude-opus-5 or openai/gpt-5.6-sol. Neither one drops the prefix.
TokenRouter offers a Claude SDK-compatible endpoint alongside the OpenAI-compatible one, with a base URL of https://api.tokenrouter.com/v1. Claude Code is the exception: for ANTHROPIC_BASE_URL you omit the /v1. The client appends /v1/messages itself, so including it doubles up the /v1 and the request fails. Even within the official site, the FAQ and the setup guide contradict each other on this.
There’s one more constraint. TokenRouter supports the CLI version of Claude Code only — the docs explicitly state that Claude Desktop, and the desktop and web versions of Claude Code, are not supported.
The migration cost itself is small. But it’s always these seams where things fall over.
The deciding question: who receives the invoice?
If you’re torn, ask who the invoice for that API key lands on.
If it lands on your own credit card, OpenRouter is plenty. Every fee is published as a number, BYOK gives you an escape hatch, and you get immediate access to 500-plus models. Within the scope of this research, I found no reason to pick TokenRouter in that position.
If it’s a company expense and people other than you share the key, what you need isn’t model selection — it’s being able to explain things after the fact. Three signals say it’s time to reconsider: when someone asks you to produce audit logs, when you have to split costs by department, and when customer data starts appearing in prompts. Whichever month you hit one of those is the month you revisit your gateway.
Before you move, though, send the same prompt to the same model on both and compare the bills. TokenRouter doesn’t disclose its markup, so the only way to know whether it’s cheaper is to run it and see. Numbers nobody publishes are numbers you have to measure yourself.
OpenRouter is already listed in the AI Deck catalog, where you can check it alongside its tags and links. If a gateway you’re curious about isn’t there yet, you can suggest it through the request feature.