Show HN: TTFT benchmark: LLM Gateway vs. OpenRouter (Claude-haiku-4.5, 150 runs)

A Show HN post titled TTFT benchmark: LLM Gateway vs. OpenRouter (Claude-haiku-4.5, 150 runs) points to a performance write-up on the LLM Gateway blog. The piece compares routing layers that sit in front of the same model, Claude-haiku-4.5, and measures time to first token (TTFT) across 150 runs. The linked article frames the test as OpenRouter vs. Vercel vs. LLM Gateway. On Hacker News the thread had 3 points and 0 comments at capture time.

TTFT is the interval from request send to the first streamed token. Holding the model fixed (Claude-haiku-4.5) and repeating the call 150 times isolates gateway and network path differences from model generation variance. The blog’s three-way setup targets products that share a similar role: a single API surface over one or more LLM backends, with OpenRouter and Vercel in the same comparison as LLM Gateway.

For builders who already proxy Claude through a gateway, TTFT is the latency users feel on chat and agent UIs before any content appears. A gateway that adds meaningful first-token delay under identical model conditions changes product feel even when completion quality is unchanged. Repeated runs (here, 150) matter more than a single timing; one-off TTFT numbers are noisy and easy to game with warm paths or lucky samples.

The market context is a crowded middle layer: OpenRouter as a multi-provider router, Vercel’s AI/gateway path in the same stack many Next.js apps already use, and LLM Gateway positioning itself with its own published numbers against both. Publishing a head-to-head TTFT study on Claude-haiku-4.5 is a direct bid for attention among teams choosing a default proxy rather than calling providers one by one.

What to do next: read the full methodology on the LLM Gateway post (region, client location, streaming settings, cold vs warm runs) before treating any ranking as settled. Re-run the same Claude-haiku-4.5 TTFT protocol from your own region and client stack if gateway choice is on the critical path. Watch whether OpenRouter or Vercel respond with matching public numbers, and whether the HN thread picks up once more people dig into the 150-run design.