Jev drops the waitlist, Theo trashes instant compaction
idan levin's WebMCP benchmark went from 25 of 49 tasks to 49 of 49 once he handed Jev website tools instead of raw page controls. That is the week in one line: Jev is strong at picking from a good option list and weak at building one.
Built with it
Ira Bodnar says Jev cut his SEO agent cost by 90%
Ira Bodnar said agents that audit and fix a client's SEO and GEO used to cost about $250 and now cost 90% less. He credits roughly 30x faster reads across nine steps, including Search Console and PostHog data and scans of who ChatGPT and Claude cite.
Jev was adopted faster than any other model in AI Gateway history.
Someone measured it
Kyle Jeong's Stagehand branch: Act latency 1.97s down to 0.46s
Kyle Jeong wired Jev into Stagehand as a decision layer, picking which element to click or field to fill while deterministic code executes. Early evals cut median Act latency from 1.97s to 0.46s and moved success from 85% to 95.4% across 240 real-site task runs. Uncertain choices fall back to an LLM.
Extract is especially interesting because Jev can’t generate text.
idan levin: Jev plus Mercury solved 49 of 49 WebMCP tasks
idan levin's open WebMCP benchmark ended with Jev plus Mercury 2.5 solving 49 of 49 tasks at roughly 112x lower model cost than GPT-6 Astra using computer use with code execution. On the same harness without WebMCP tools, Jev solved 25 of 49. Separately, Guillermo Rauch said Jev runs up to 18x faster at p95 and more accurate than GPT Luna in fx's safety reviewer.
Jev’s browser-control accuracy on its own was not amazing
Writing criteria
tamara and Zachi skip the summarization prompt, Jev scores instead
Two builders landed on the same shape this week: instead of a summarization prompt, Jev scores each candidate directly. tamara's compaction script scores every tool call and drops what's irrelevant rather than summarizing, and Zachi's jev() Postgres extension judges rows one by one with no index and no embeddings, 129 rows in about 1s for $0.0009 and 6ms from cache on the rerun.
Compaction isn't a filter
The case against
Theo calls Jev instant compaction a terrible strategy
Theo tore into tamara's instant compaction demo, arguing that filtering history on a probability threshold is not compaction. His objections: Jev scores each tool call without seeing its result, frontier reasoning comes back as encrypted payloads it cannot read, and editing history forces a cache rewrite of everything after the edit.
if you think this style of bs filtering on a probability threshold is actually a compaction strategy, I highly recommend you just use the defaults in tools like Claude Code and Codex
Frequently asked questions
How much did Jev reduce SEO agent costs?
According to Ira Bodnar, Jev cut the cost of SEO and GEO audit agents by 90%, bringing them down from about $250 by achieving roughly 30 times faster reads across nine steps including Search Console and PostHog data.
What speed improvements did Kyle Jeong see when integrating Jev into Stagehand?
Kyle Jeong's integration cut median Act latency (the time to execute a decision) from 1.97 seconds down to 0.46 seconds and improved success rates from 85% to 95.4% across 240 real website task runs.
How did Jev perform on the WebMCP benchmark?
Jev plus Mercury 2.5 solved all 49 WebMCP tasks at roughly 112 times lower model cost than GPT-6 Astra, though Jev alone solved only 25 of 49 tasks without the Mercury tools.
What new approach did builders use instead of summarization prompts?
Instead of writing summaries, builders like tamara and Zachi are using Jev to score each candidate or row directly, allowing them to drop irrelevant items without creating summaries.
What was Theo's main criticism of instant compaction with Jev?
Theo argued that filtering history on a probability threshold is not true compaction, and that Jev scores tool calls without seeing their results and cannot read encrypted payloads from frontier reasoning.