Jev Daily

OpenRouter clocks GPT-6 Luna at 180ms, Jev second

The only number OpenRouter disclosed in the post belongs to the model that won.

OpenRouter published a latency ranking for decisions models and put GPT-6 Luna in first place at 180ms on global requests. Jev comes second and Perplexity Decider third. Those are the only three models named in the post.

OpenRouter
@OpenRouter
X
GPT-6 Luna is the fastest Decisions model today, at 180ms on global requests
Oct 8, 2026 · View on X

The ordering is the entire disclosure. OpenRouter did not put Jev's latency in the text, so the gap between first and second is unknown. It could be ten milliseconds or it could be two hundred. Anyone treating this as a reason to switch is reading a number that is not there.

What the 180ms does and does not cover

OpenRouter describes the figure as applying to global requests, which is a mix of traffic across regions and callers rather than a controlled test on a fixed workload. That matters for a latency comparison more than it does for an accuracy one. A model with a different geographic footprint, a different share of large payloads or a different burst profile is not being measured on the same inputs as its neighbours in the list. The ranking tells you what OpenRouter's users experienced. It does not tell you what your own calls would do.

It is also worth separating this from the other OpenRouter numbers already in circulation. The company previously reported that Jev took 27% of its classification requests, which is a volume share, and separately that Solar Decide beat Jev on JevBench, which is a benchmark result. A speed ranking, a traffic share and a benchmark score are three different claims, and none of them substitutes for the others.

For people building on it

If latency is the thing you are optimising, this post is a pointer, not evidence. The usable takeaway is that at least one provider's aggregate traffic has Luna ahead of Jev on response time, and that second place is still second place even when the margin is undisclosed. Measure on your own request mix, from your own region, before you move anything. The board will not do that for you.

Get the next one by email

Jev, read daily so you do not have to. The builds, the benchmarks, the criteria that worked and the cases where it lost, from the people shipping on TypeSafe AI's System One model.