Published
Updated
One screenshot decision was $0.0002 on Jev and $0.032 on Opus 5
This topic was created 31 days ago, and the information it contains may have evolved or changed since then.
awlevin's macOS loop reads the screen with OCR and asks Jev which action comes next. On one shared screenshot the README prints $0.0002 and 0.13 to 0.38 seconds for Jev, against $0.032 and 5.2 seconds for Claude Opus 5 on the bare image. The same page says Opus read the dates off the pixels, and this loop annotates dates in code before the classifier runs.
Somi posted on September 23 that awlevin/typesafe-computer-use does a computer-use step for $0.0002, against $0.032 for Claude Opus 5 on the same screenshot. The trick in the post is OCR, then Jev picks the click. The catch, also in the post: Opus reads dates off the pixels, and this project parses them in code. The README is MIT. It targets macOS 14 or newer. Windows is marked experimental, and the page says that adapter has not been run on Windows.
The priced row is one decision, the same screenshot and the same goal. Jev: 4,882 input tokens, $0.0002, model latency 0.13 to 0.38 seconds, about 1.5 seconds end to end once capture and OCR are included. Opus 5 on the bare screenshot: 4,785 input tokens, $0.032, 5.2 seconds, about 5.5 seconds end to end. The README calls that 155 times cheaper, 14 to 40 times faster for the model call, and 3.7 times faster end to end. A loop that keeps history is still $0.0002 on the Jev side and $0.035 to $0.08 on the Opus side. A 12-step task is $0.003 against $0.40 to $0.90. The page’s caveat is the date one Somi quoted. Anything the frontier model does from the pixels has to be rebuilt here as state. dates.py stamps “dated 2026-10-13 (in 27 days)” onto a block that contains a date, and says “near a line dated …” on its neighbors.
The desktop path captures the frontmost window, runs Vision OCR, walks the accessibility tree, and sends one TypeSafe request with three Choice questions: the kind of action, which item, and which site. A writer model, defaulting to Claude Haiku for text and Sonnet for the final read, is called when a field needs free text or the classifier has stopped. The loop stops on done, on confidence under 0.4, after a stall, or at the step cap. --act moves the real mouse. A dry run does not.
A second backend never looks at pixels. It reads the DOM over the Chrome DevTools Protocol, on a temporary Chrome profile. Perception, same page, same moment: a local fixture is 1.3 ms and 28 elements against OCR at 288.0 ms and 31 blocks. Hacker News is 4.5 ms and 120 elements against 697.3 ms and 46 blocks. The Singapore Wikipedia page is 14.4 ms and 91 elements against 897.0 ms and 50 blocks. OCR matched 7 of 13, 7 of 88, and 3 of 85 DOM labels verbatim. The end-to-end browser loop is 302 to 380 ms a step at the median, and most of that is the TypeSafe call. Canvas and text baked into images are invisible to the DOM path. We did not run either loop.
Browser Use’s flights demo and Stagehand already start from a list of browser actions. This README prices one desktop decision against Opus 5 and does not print how often the goal was reached. dejevu times a finished flight search. That is a different clock.
This site's reading
Editorial notes evaluating claims against primary sources, contextualizing findings alongside related implementations, and defining technical terms.
Verify
Primary post is @somi_ai, September 23, 2026, 04:40 UTC. Display name Somi. Text: computer use at $0.0002 a step, Opus 5 on the same screenshot at $0.032. awlevin's approach is OCR, then Jev picks the next click.
Opus reads dates off pixels. Here the date parse is code. The post says 155x and links the repository. README: MIT, macOS 14 or newer, Windows experimental and untested on Windows.
One decision, same screenshot and goal. Jev: 4,882 input tokens, $0.0002, model latency 0.13 to 0.38 s, end to end about 1.5 s. Claude Opus 5, bare screenshot: 4,785 input tokens, $0.032, 5.2 s, end to end about 5.5 s.
The page calls that 155x cheaper, 14x to 40x faster on the model, 3.7x faster end to end. With history the Opus cell is $0.035 to $0.08. A 12-step task is $0.003 against $0.40 to $0.90.
The caveat on the page: the big model read event dates off the pixels. The classifier needs dates.py. DOM versus OCR, one decision each: local fixture 1.3 ms (28 elements) against 288.0 ms (31 blocks); Hacker News 4.5 ms (120) against 697.3 ms (46); Singapore Wikipedia 14.4 ms (91) against 897.0 ms (50).
OCR reproduced 7/13, 7/88, and 3/85 of the DOM labels verbatim. Browser-backend loop p50 is 302 to 380 ms a step. Default stop under confidence 0.4. Drives the real mouse in --act.
We did not run it.
Compare
Browser Use's flights demo and Stagehand are click loops that already assume a browser action list. This README prices one desktop decision against Opus 5 on the same screenshot. It does not print a task success rate.
dejevu's flights table is a finished search, 5.63 seconds against a published Jev cell of 7.09 seconds, which is a different measurement from a per-decision price.
Terms
- one shared screenshot
- The cost table in awlevin's README. One decision, same screenshot and goal. Jev is $0.0002 and 0.13 to 0.38 seconds. Claude Opus 5 on the bare image is $0.032 and 5.2 seconds. The page says Opus read the dates from the pixels and this loop annotates them in code.
- OCR label overlap
- On three pages, OCR reproduced 7 of 13, 7 of 88, and 3 of 85 DOM labels verbatim. The README's reading is that a classifier over OCR blocks is choosing among garbled strings.