Cloud baseline

GLM 5.3 FlashX — 0.6303 on Gauntlet 7.2

z-ai/glm-5.3-flashx · scorer sb-7.2

GLM 5.3 FlashX, single model via OpenRouterz-ai/glm-5.3-flashx · single agentOct 4, 2026
Overall
0.6303
Excellent
no
Benchmark
Gauntlet 7.2
Model build
49m 53s
Prompt tok
25.1M
Gen tok
226.9k

Recorded inference cost

At least $2.62

From the saved OpenRouter generation billing summary (142 recorded requests). This is a lower bound; the billing record is incomplete. Includes auxiliary model calls made during the build. Scoring reuses the saved build; this amount describes the original inference spend.

  • The harness counted tokens no recovered generation id accounts for: prompt -972, completion -5

Token usage

Tokens: 25.1M in · 227k out · 24.8M cached (exact, OpenRouter records)

25,300,894 tokens over 142 model calls: every call of the run, the model's own and the harness's helper calls, counted from OpenRouter records. Reasoning tokens as the provider reports them: 129,086 (not added to the total).

  • z-ai/glm-5.3-flashx-20260918

    In 25.1M · Out 227k · Cached 24.8M · Reasoning 129k · 141 calls

  • google/gemini-2.5-flash

    In 972 · Out 5 · Cached 0 · Reasoning 0 · 1 call

Tier breakdown

A · structure & runtime1.00
B · behaviour1.00
C · vendor contract0.95
D · finesse0.93

Scoring detail

How this score was built

Per-check scorer output, as posted by the app. Expand a section to inspect its checks.

Earned credit before admission

(0.88 × 0.8614 core + 0.12 × 0.99 gate × 1.00 excellence) × 0.9429 critical = 0.8268

Critical defects compound a multiplier on the whole score (pre-severity 0.8769):

j_workflow_journey 0.86 → ×0.94

The core (88% of the total) is the weighted mean of the thirteen measured tiers. The last 12% is the excellence slice: it unlocks in proportion to the perfection conditions below (15 of 16 met here), then pays out at the excellence tier's own measured mean. Core 0.7580 + excellence 0.1189.

j_first_use 1.00j_workflow_journey 0.86j_error_state 1.00j_empty_state 1.00console_clean 0.00v_responsive_375 1.00v_dates_readable 1.00t_scene_binding 1.00x_conservation_residual 1.00r_no_row_loss 1.00p_drag_frames 1.00p_idle_flatness 1.00p_stream_apply 1.00p_under_stream 1.00p_api_latency 1.00p_sync_wall 1.00

Earned score 0.827 · Admission ceiling 0.639 · Final score 0.630

Final score is the lower of earned credit and the admission ceiling. Passing admission adds no points.

Visible: passedMatching: not metGood: not metVisual excellence: not metBackend excellence: passed

Tower structure, currency mapping and scene truth: s_tower_geometry, t_height_pixels, t_vs7dbg_truth (maximum 0.639)

3D interaction and overview legibility: t_camera_math, t_labels_culling, t_brush_link, t_stream_diff, q_overview_legibility, q_inspector_framing (maximum 0.699)

Event animation and backend recovery: m_committed_event_replay (maximum 0.899)

Run notes and corrections

Scoring note, 2 November 2026. This run was graded by the sb-7.2 scorer as it was before harness version 3.0.148. That scorer only read notification text from elements with nothing nested inside them, unless the feed used a class name containing "notif". This build puts the kind in its own element inside each entry, so the scorer read none of its notifications and couldn't tell whether one appeared after the approve and reject steps. The notification parts of j_workflow_journey and j_workflow_reject may therefore understate how the model did. This build's files weren't kept, so it can't be re-graded with the fixed scorer, and the score stays as published.

Token rates

Measured by the engine itself, one record per completed model call: prefill rate is prompt tokens over time-to-first-token, generation rate is completion tokens over the decode window. Medians per node.

NodeCallsPrompt tokGen tokPrefill tok/sGen tok/s
gemini19725909.3138.9
z14125,073,026226,89147422.5140.6
fleet14225,073,998226,89647403.3140.6

Graded browser recording

Loading duration…

Watch the full graded browser recording at its original speed: the payment field, structure inspection, committed updates and replay. The check results and screenshots on this page provide the wider evidence. Pause or scrub to inspect a view.

Each tower represents a payment: height encodes its amount, cap color its status, and the collar its currency. The selected detailed tower is expected to animate its collar after a committed backend update or replay. These are the task requirements; the check results show what this app actually achieved.

Ready to play0:00 / —:—
Open recording file

Drag the timeline to seek. Keyboard: arrow keys move 5 seconds; Home and End jump to the start and end.

Graded browser recording re-encoded to 766x478 at 25 fps to fit 4 MiB, same 124.6 s; graded original sha256 f5e7002fb1d234496a589bd661f0d2118e565750596ab63fced2e2040dcdaf71

Screenshots

Captured from the built application during browser grading. Captions identify the recorded views and checks; these images do not represent repair rounds.

1Initial app view
2Error state
3Payment workflow
4First captured render
5Latest captured render

Run details

Model
z-ai/glm-5.3-flashx
Engine events
0
Repair rounds
0
Started
Oct 3, 2026, 02:57 AM
Finished
Oct 3, 2026, 03:55 AM
Re-scored
Oct 4, 2026, 05:27 AM (same build, corrected scorer)

How the benchmark is scored