Lightpanda times Jev at 0.53s on login fields, Haiku at 7.83s

Lightpanda has taught its agent to log in to websites, and it is using Jev to do the field picking. The company posted that Jev now handles logins in its agent, identifying the username field, the password field and the submit button from the page.

Lightpanda

@lightpanda_io

X

Jev now handles logins in Lightpanda's agent. It picks the username field, password field and submit button from the page.

The numbers it posted are 0.53 seconds and 2 calls for Jev, against 7.83 seconds and 8 tool calls for Claude Haiku 5.5 with browser tools. That is roughly fifteen times the wall clock and four times the calls for the Haiku path on Lightpanda's own setup.

What the number does not cover

Lightpanda did not say how many pages it tested, which sites, or whether the logins actually succeeded. A login form is a narrow task with a clear right answer, so a single-shot classifier against an agent loop with browser tools is close to a best case for the classifier. The gap in tool calls is the part that explains the gap in seconds. Two calls versus eight is the difference between reading the page once and poking at it until something looks like a form.

There is also the matter of who ran it. Lightpanda already ships Jev inside its binary, which it used earlier to separate real page content from cookie walls and consent screens. This is the integrator timing the stack it chose, not a neutral comparison, and the post reads like one. TypeSafe AI, which makes Jev, replied to it.

For people building on it

If you are building browser automation, the useful read here is the call count rather than the latency. An agent that resolves a form in two calls is cheaper and more predictable than one that resolves it in eight, and the seconds follow from that. Whether it generalizes is unanswered. Login pages vary wildly, some are multi step, some hide the password field until the username is submitted, and none of that is in the post.

The honest summary is that one integrator with a reason to like the result measured a fifteenfold speedup on a task it designed around. That is worth knowing and it is not a benchmark. If logins are on your critical path, the test to run is your own list of sites, with a success rate next to the latency, because a fast wrong field selection costs more than a slow right one.