Every Strata build with a complete webbench, against Sonnet 5.5

• Newton (Gyro-L) is the strongest Strata build: it passes 120 of 134 webbench tasks strictly (mean 0.95), 9 more than Galilei (Gyro-M) at 111 (0.92), then Kepler (Swift IQ3_XXS) 107 (0.91), Euler (Flash-Next IQ2_XS) 88 (0.80) and Turing (Flash-Next Coder IQ1_M) 32 (0.38). Newton is also the only Strata build that built all 134 projects.
• Most of Newton's lead over Galilei is on the first 100 tasks (89 against 82); on the 34 React and Next.js tasks it leads 31 to 29, and on Next.js alone 17 to 16. Five tasks are passed by Newton and by no other Strata build.
• The cost is speed. Newton writes as much as Galilei (a median of about 25,800 output tokens per task against 25,700, most of it thinking) but decodes at 32 tok/s median against Galilei's 68, so a task takes a median 788 s against 380 s, and the whole webbench 27.1 hours against 13.1. The grill shows the same gap: 49 minutes of task time per run against Galilei's 23.
• On the canonical grill Newton scores 22 of 23 in all three runs and completes the long-horizon agent round 6 of 6 each time, the same as Sonnet. It is the only Strata build that matches Sonnet on both. Euler also scores 22 in every grill run but completes the agent round 5, 6 and 6 times; Galilei scores 21, 22 and 22 with 6 of 6 every run. On vision Newton scores 49, 50 and 50 of 52; Galilei's best run is the only perfect 52 of 52 on this page.
• Turing is pruned to half its experts for code: it scores 20 to 22 on the grill and 49 on vision in every run, but builds only 95 of the 134 web projects and passes none of the 20 Next.js tasks.
• Each Strata build ran one webbench sample, at different sampling depending on the engine version (see the notes under the scoreboard), so a gap of a few tasks between two builds may not hold up on a second sample.
• Sonnet 5.5 is still ahead at 129 of 134 with a 0.99 mean, and it finishes each task in about 18 s. Its times and token counts come through the API proxy, so they aren't comparable with the local runs.

Scoreboard

Every local model ran one webbench sample with no temperature sent, so each uses its server's default, with reasoning effort set to xhigh, a 24k-token thinking budget and a 32k output cap. Sonnet's column is its first of three samples; its times and token counts come through the API proxy and aren't comparable with the local runs.

Sampling differs between Strata versions. With no temperature sent, engine 0.1.40 always picks the most likely token, while engines 0.1.47 (Strata 0.2.2) and 0.1.52 (Strata 0.2.7) apply the model card's sampling: temperature 1.0, top_p 0.95, top_k 20. Kepler and Turing ran on 0.1.40, Galilei on 0.1.47 and Newton on 0.1.52. Euler answered webbench tasks A-001 to A-065 (62 tasks) on 0.1.40 and the other 72 tasks, the grill and the vision test on 0.1.47.

Newton's grill and vision runs (2026-10-31) used the same settings as Galilei's: no temperature sent, reasoning effort xhigh. Turing's grill and vision runs are older (2026-10-06) and ran at temperature 0.2 and 0.1; the other Strata grills sent no temperature. Galilei's webbench was paused for 14 minutes during task A-029, so that task's wall time and the suite's wall hours include the pause.

Webbench by category

Strict passes out of the tasks in each category. Bars are drawn to the same 0–100% scale, in the page's model order.

Where the Strata builds differ

Counted over the 134 webbench tasks, all of which every Strata build here has finished.

How each model fails on webbench

Canonical grill, task by task

Each square is one run, three runs per model. Highlighted rows are the tasks where the models differ.

expr_eval rejects inputs like 1++2; the local models miss that edge case almost every run, as nearly every model tested here does.

All 134 webbench tasks

Score is the functional score of the first sample (1.00 = every check passes). Green is a strict pass, amber a near miss at 0.80 or above, red means nothing built. Sonnet runs: ● pass, ○ fail, for each of its three samples.