Sentry screens 10M events a day on Jev for $263

Sentry runs Jev across 10 million events a day to screen for prompt injection before a coding agent reads an error report, and Cloudflare Developers says that costs $263 a day. The same post says the team is now testing its own classifier, aiming to match Jev's accuracy for 90% less. @grichadev explains how in the linked write-up.

Cloudflare Developers

@CloudflareDev

X

The next prompt injection your coding agent reads might be an error report.

Cloudflare Developers

@CloudflareDev

X

Now the team is testing its own classifier, aiming to match Jev's accuracy for 90% less.

The per-event number

Arithmetic on the two posted figures puts the cost at roughly $0.0000263 per event, or about 2.6 cents per thousand events screened. That is the kind of number worth carrying around if you are deciding whether to put a model in front of untrusted text at ingest volume rather than sampling it. At 10 million events a day, a screening pass that most teams would assume is prohibitive lands near the cost of a mid tier SaaS seat.

The threat model is the part that generalizes. An error report is attacker influenced text that a developer tool reads without much suspicion, and a coding agent that pulls in a stack trace or a user supplied message is reading whatever got put there. Screening at the ingest boundary means the agent never sees unchecked content, rather than relying on the agent to recognize an instruction aimed at it.

The classifier is a target, not a result

The 90% figure is a goal the team has set for its own classifier, not a measured outcome. Cloudflare Developers describes it as testing, and nobody has posted whether the in-house model matches Jev's accuracy at 10 million events a day, what accuracy Jev itself hits on this traffic, or what the false positive rate looks like when a legitimate error report gets flagged as an injection attempt.

For people building on it, treat the $263 as the reference point and the 90% as an unverified aspiration. A screening classifier that is cheaper and slightly worse is a different product than one that is cheaper and equivalent, and the difference only shows up in the eval nobody has published yet.