Why it mattersA team choosing a coding harness this quarter has evidence that the harness can matter as much as the model on a real front-end task, so any benchmark that names only the model is missing half the information.
A developer publishing as Alvin S has posted timings for one prompt run against ten combinations of coding models and coding harnesses. The post reached the Hacker News front page with 95 points and comments in the last eight hours. The page carries no dated byline, and the numbers reflect one developer's setup on one machine, so treat them as one data point.
The prompt asked each combination to output a self-contained single-page Three.js sci-fi hangar with hovering drones, animated warning lights, emissive runway strips, subtle volumetric-style fog planes, a drone-formation toggle, and a cinematic camera path.
The times, on the same task
The fastest run was Qwen 3.8 27B x-high on OpenCode, at 8 minutes 48 seconds. The slowest was the same model on OMP, at 41 minutes 25 seconds. The same model on OpenCode used 707,307 total tokens; the OMP run used 3,479,386, close to five times as many. Astra 6.0 Max on Codex ran the same prompt in 37 minutes 29 seconds. GLM 5.3 Flash Max finished in 9 minutes on Codex, 20 minutes on OpenCode, and 30 minutes on OMP.
The Codex harness looked fast on this prompt overall: GLM 5.3 Flash Max in 9 minutes, Luna 5.6 Max in 9 minutes 13 seconds, SOL 5.6 Max in 10 minutes 48 seconds, and Astra 6.0 Max in 37 minutes 29 seconds. That last figure is the interesting one: the harness held constant, the model changed, and the wall time rose by a factor of 3.5.
Cache-hit rate was high for every run and did not track wall time. GLM 5.3 Flash Max on OpenCode had the highest cached input at 96.89 percent, and it still took 20 minutes 28 seconds. So a high cache-hit rate on its own does not predict which combination will finish first.
The output check
Only some of the runs produced something that actually rendered. Seven of the ten runs opened in a browser, six of them passed a screenshot check, and one Codex run was recorded as "blocked" at the browser step. Tool-error counts ranged from zero (two OMP runs and two OpenCode runs) to 89 (Qwen 3.8 27B x-high on OMP), on the same prompt, which is the number that separates the harnesses most.
DeepSeek Harness, both stock and in the PTC configuration, produced results in 18 minutes 15 seconds and 24 minutes 32 seconds respectively on Qwen 3.8 27B x-high, with the stock DSH run finishing without a screenshot pass and the PTC run passing.
The reason to read a page like this is the one thing that does not fit on a leaderboard: a model published a benchmark of itself, and the number that gets quoted is often the best case in the best harness. This table shows the range for one task and one prompt, so a team picking a harness for the next front-end refactor can start with a shorter list.
Source
This item was written by an AI system from the linked source. Reveneau is responsible for what it publishes.
Get AI News in your inbox
New developer tools, model and agent releases, and how teams are actually using them to release software. Short, and only when there is something worth reading.
