Jev Daily

Browser Use lifts cache hits to 96%, task cost falls to $0.45

A harness fix, not a capability gain, and the score did not move.

Browser Use says its last eval surfaced a cache hit rate of 78%, which it treated as a harness problem rather than a model one. After the fix, the team reports 96%, cost per task down 33% from $0.67 to $0.45, and run time down from 5.0 to 4.6 minutes. The score stayed the same.

Browser Use
@browser_use
X
our cache hit rate was kinda low (78%), so we fixed the harness
Nov 13, 2026 · View on X

Co-founder Gregor Zunic posted the same two numbers, 78% to 96%, on the same day. The change ships in browser-use 1.0.4, installable with `uv add "browser-use==1.0.4"`.

What actually moved

The cache hit rate here is the share of prompt tokens served from the provider's prompt cache instead of being billed and processed fresh. A low hit rate usually means something in the prompt prefix is churning between steps, so each turn re-pays for context the model has already seen. On a browser agent that takes many turns per task, that cost compounds quickly, which is why an 18 point swing in cache hits shows up as a third off the bill and a few seconds off each run.

The flat score is the important part of the claim. Browser Use is not saying the agent got better at tasks, it is saying the same behaviour now costs less to produce. That is a plumbing win, and plumbing wins are the kind that hold up, but they also cap out. Once you are at 96% there is not another 33% waiting behind the same lever.

For people building on it

These are the vendor's own numbers on the vendor's own eval, with no task list or harness detail published alongside the post. The $0.67 and $0.45 figures are averages over whatever mix of tasks that eval contains, so your per-task cost will track your own page complexity and step count, not theirs. What transfers is the diagnosis rather than the dollar figure. If you are running a multi-step agent and have never looked at your cached token share, the gap between 78% and 96% is a reminder that the number is worth checking before you blame the model for your bill.

The release follows Browser Use's earlier 1.0.0 rewrite, and lands as a point release on that line.

Gregor Zunic
@gregpr07
X
We improved caching rate from 78% -> 96% on browser use
Nov 13, 2026 · View on X

Know what Jev is good at before you bet on it

One email a day. Under three minutes. What builders found out about Jev yesterday.

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.