dejevu books the same Flights search faster than the published Jev demo, on another machine

Idov Mamane posted idovmamane/dejevu on September 22. The post says Browser Use plus Jev booked Google Flights in 7.1 seconds, and dejevu plus plain Llama 3.3 70B did it in 5.6 seconds, with 10 model calls against 17 and 5.6 times fewer tokens. Traces are in the repo. The license is MIT.

The task is the one from browser-use/jev-ultrafast: one-way Zurich to London on a Sunday, one adult, economy, stop when results are visible. dejevu asks an ordinary chat model for one JSON object, op, target, text, option, and key. The model names an element number. The page resolves the node. A separate text model is not in the loop. Done is accepted only on a settled page, and code checks the final page. The README says the model saying it is finished does not count.

The flights table is a median of verified runs. jev-ultrafast with Jev 1.13 plus Mercury 2.5, marked published: 3 of 3, 7.09 seconds, 17 calls, 13 actions, 84,650 input tokens, about $0.0036 at list price. dejevu with llama-3.3-70b on Groq: 3 of 3, 5.63 seconds, 10 calls, 9 actions, 15,071 input tokens, $0.0092. gpt-oss-120b on Cerebras passed 1 of 2 at 7.92 seconds. The same Llama on DeepInfra passed 2 of 2 at 37.5 seconds and $0.0020. gpt-oss-20b on Groq passed 1 of 4. gemini-2.5-flash, gemini-2.5-flash-lite, qwen3-next-80b, deepseek-v3.1, and gpt-4.1-nano show zero verified flights runs, mostly by stopping inside the calendar or looping.

Wikipedia, from the main page to the Gödel incompleteness article, is a second table. The published jev-ultrafast cell is 2.80 seconds, and the verified count is blank. llama-3.3-70b on Groq is 3 of 3 at 1.31 seconds, two calls, one action, 4,033 tokens, $0.0024. Several other chat models also finish that task in two calls. Three hold-out tasks, added after the harness was tuned, are described as 18 verified runs out of 18 across three models and two runs each.

Three clocks sit next to each other in the repo. The table median is 5.63 seconds against 7.09. A side-by-side video freezes the counters at 7073 ms and 6070 ms, using jev-ultrafast’s docs/demo.mp4 on the left, cropped, with a counter overlay. A GIF caption says 6.07 seconds. The README says the Jev numbers were measured on that project’s machine and Chrome profile. dejevu’s runs are from September 22 on a MacBook Pro M3 Pro, headless Chrome 153, through OpenRouter. September 20 has passed, so the search date in this repo is Sunday, October 18, 2026. A fresh profile hits Google’s consent page first. The agent dismisses it, and the clock starts on the Flights page, which the README says is the same boundary jev-ultrafast uses.

The README also says what the speed does to the bill. dejevu sends about 5.6 times fewer tokens. Jev’s list price in that note is $0.042 per million. Groq at $0.59 per million makes the Flights run about 2.5 times Jev’s price. The break-even it prints is a provider at $0.24 per million for the same 70B model. DeepInfra at $0.10 per million was cheaper than Jev on the two passing runs and took 37.5 seconds. TypeSafePolicy can replay the Jev arrangement on the same machine with --backend typesafe. The README says that path is untested because no TypeSafe key was available.

Offline tests in the README: 45 pytest cases with no browser and no paid calls, and 18 browser checks with no model calls. The limits section says three runs per cell show a difference and are not a benchmark, the harness was tuned on llama-3.3-70b, and two websites plus a local fixture do not establish general reliability. We did not run it.

The flights story on this desk is the earlier clip: about seven seconds and $0.0039, with Jev choosing the action and the target. dejevu’s comparison uses the published side of that demo and a new date, a new machine, and a chat model that returns the whole action in one object.