226 projects tested, and "Node.js 18.19.1 is too old for Next.js 16 (requires >=20.9.0)" stopped 4: week of September 7, 2026
226 projects installed and run on clean machines in the week of September 7, 2026, 229 verified against the real thing. Every number links to a recorded run.
Covering 2026-09-07 to 2026-09-14. Published .
What was measured
- Projects run
- 226 over 316 attempts (263 the week before)
- Verified against the real thing
- 229
- Verified with stand-ins only
- 30
- Could not be verified
- 57
- Tested here for the first time
- 199
- What it cost us
- 44.78 USD, 55% of it model tokens
What it ran on
Argusic publishes other people's server prices every day. These are its own: the machines this week's work was rented on, at the rates actually paid, each one destroyed when its job finished.
- Attempts on CPU machines
- 318 at 0.1186 USD an hour
- Attempts on GPU machines
- 1 at 0.7400 USD an hour
- Data centers used
- EU-CZ-1, EU-NL-1, EU-RO-1, EUR-IS-1, US-CA-2, US-KS-2, US-MD-1, US-MO-1, US-MO-2, US-NC-1, US-NE-1
What the week cost, broken down
A single total answers nothing for somebody working out whether they could do this themselves. The difference between the dearest and the cheapest project is how much fixing each needed, not a price anybody chose.
What the pool looks like now
The state of the whole body of evidence at the end of this week, not only what the week itself added.
- Projects measured in total
- 470
- Of those, reached a working state
- 342
- Most measured languages
- TypeScript (116), Python (108), Rust (65), Go (52), JavaScript (41)
What kept breaking
Counted across projects, not across attempts, so one repository retried three times is one problem. Each label is the wording an agent actually used.

- 4 projects
- Node.js 18.19.1 is too old for Next.js 16 (requires >=20.9.0)
- 3 projects
- No JDK installed in container
- 3 projects
- pnpm not available in PATH
- 3 projects
- pnpm not installed
- 2 projects
- @tailwindcss/oxide missing native binding for linux-x64-gnu
The result most worth reading
cubrid scored 100 of 100, after hitting 11 errors that the agent fixed with no human help.
CUBRID 11.5.0 debug build compiles and installs successfully with all core binaries (cub_server, csql, cub_master, cub_broker, cub_admin, cub_pl, cub_cas, loadjava, cubrid_rel) and 3 unit test binaries functional.
The agent's own words, from its recorded report.
What this week does not tell you
- Behavior over time
- each attempt uses the smallest documented settings and a fixed time budget, so this measures whether a thing runs, not how it holds up under load
- Anything not tested
- only the 226 projects above were run this week; a project absent from this report was not measured, not judged
- Real-service behavior for 30 of them
- those ran against stand-in services, never the real ones
The report as written
Argusic installed and ran 226 projects on clean machines between 2026-09-07 and 2026-09-14, over 316 attempts, 37 fewer subjects than the week before.
229 were verified against the real thing, 30 only against stand-in services, and 57 could not be verified at all.
199 projects were tested here for the first time.
The same problems keep appearing across unrelated projects. The most common this week was "Node.js 18.19.1 is too old for Next.js 16 (requires >=20.9.0)", which stopped 4 projects. Also recurring: "No JDK installed in container" (3), "pnpm not available in PATH" (3), "pnpm not installed" (3), "@tailwindcss/oxide missing native binding for linux-x64-gnu" (2).
The result most worth reading was cubrid, which scored 100 of 100 after hitting 11 errors that the agent fixed with no human help. In the agent's own words: CUBRID 11.5.0 debug build compiles and installs successfully with all core binaries (cub_server, csql, cub_master, cub_broker, cub_admin, cub_pl, cub_cas, loadjava, cubrid_rel) and 3 unit test binaries functional.
Running all of this cost 44.78 USD, of which 55% was model tokens rather than machines.
24.43 USD of that was model tokens and 20.35 USD was machines, which is 0.20 USD for each project measured.
The dearest project of the week was pixijs at 1.14 USD, the cheapest express at 0.01 USD. The difference is how much fixing each one needed, not a price we chose.
The work ran on machines rented by the hour and destroyed afterwards: 318 attempts on CPU machines at 0.1186 USD an hour, and 1 attempt on GPU machines at 0.7400 USD an hour. Those are the rates Argusic actually paid, not list prices. The pods came up in 11 data centers: EU-CZ-1, EU-NL-1, EU-RO-1, EUR-IS-1, US-CA-2, US-KS-2, US-MD-1, US-MO-1, US-MO-2, US-NC-1, US-NE-1.
Across everything measured so far, 470 projects have been installed and run here, and 342 of them reached a working state. Most of them are written in TypeScript (116), Python (108), Rust (65), Go (52), JavaScript (41).
Every number above comes from a recorded run, and each one has its log and its terminal recording on this site.