Blog
OpenRouter vs going direct to the provider
The purist take says skip the middleman. The measured take is more interesting: the hop costs 68ms at the median, wins at p95, and the 5.5% buys something real until it doesn't.
Going direct means provider list prices with no fee and no hop, but one integration per provider. OpenRouter adds 5.5% on credits and a measured +68ms median hop (its p95 beat direct in our runs) in exchange for one API across 400+ models. Direct wins for single-provider production; the router wins for breadth.
What direct actually buys you
No fee, no hop, first-party features the day they ship, and your provider relationship (rate limits, negotiated terms, support) held directly. If your production traffic runs on one or two models and selection is settled, direct is not a compromise; it is the correct architecture, and no honest gateway comparison should talk you out of it.
The cost of direct is plumbing multiplied by providers: separate keys, separate billing, separate rate-limit behavior, separate failure modes, and a code change every time you want to try someone new.
What the hop costs, measured
We ran twenty streamed requests direct and twenty through a full routing gateway, same prompt, same hour. The gateway added 68ms at the median time-to-first-token, and its p95 was better than direct's (0.821s versus 0.925s) because provider-side variance is bigger than the detour. Latency is a measured non-issue for most traffic; anyone telling you the hop is disqualifying should show you their numbers.
The fee is the real cost: 5.5% on credits, which is cheap below roughly $900 a month of spend and a compounding tax above it.
The gateway hop, measured: 20 streamed runs each way
Time-to-first-token, same prompt, same hour: direct provider call vs a full routing gateway. Method and runnable script in the gateway comparison. Measured 2026-07-28.
The switching-cost asymmetry
Here is the part the purist take misses: the OpenAI-compatible surface means going through a router first costs you almost nothing architecturally. You can start on a router for breadth, settle your model choice with real traffic, and go direct later with a base URL change. Starting direct and adding breadth later is the expensive direction.
The unresolved question either way is selection. Direct pins you to a provider by inertia; a catalog pins the choice on you by design. If you want the frontier without owning the choosing, that is the blended-model category: Pareto, several LLMs per request, one answer, billed at cost, receipts here.
Questions, answered
Is it cheaper to use OpenAI or Anthropic directly?
Per token, prices are identical; aggregators pass list rates through. Direct avoids the 5.5% credit fee, so pure cost favors direct, weighed against maintaining one integration per provider.
How much latency does OpenRouter add?
+68ms at the median in our measurement, with p95 better through the gateway than direct (0.821s vs 0.925s). Provider-side variance dominates the hop.
Should production traffic go through a router?
If you run one settled model, direct is right. If you run several models, compare often, or want fallbacks without code, the router earns its fee, at least until spend clears the ~$900/month crossover, where flat-fee and at-cost structures deserve the comparison.
Do I lose provider features through a gateway?
Sometimes, temporarily: first-party features can lag behind direct APIs on any middle layer. Check the specific features you rely on (caching semantics, tool calling, streaming details) before committing either way.
Compare any two models
Rates verified 2026-07-28. "Measured task" = our identical dashboard-generation prompt, metered where marked ✓ and list-math otherwise. Verbosity from the Verbosity Index, Edition 1. Data: prices.json.