Benchmarks Studio · job receipts

What the Studio's own jobs did, not how the pictures looked.

86 real jobs from the Media Studio production table, recorded between 13 Nov 2026 and 15 Nov 2026. Each route's saved rate, saved-job latency and recorded cost estimate come from the row itself. No prompt, project, asset or media key appears in this page or its data file.

Jobs in the window
86
Saved
66 (76.7%)
Routes and kinds
17
Checked
16 Nov 2026, 05:55 UTC
Freshness
Stale · 5 days old

This is reliability and latency evidence, not a quality ranking. The separate matched-still comparison has 3 recorded preference submissions (1 blind, 2 unblinded, 0 reveals without choosing). Those are 3 unauthenticated observations between two saved clips, not a quality ranking of these 17 routes.

This receipt is older than its own 2-day limit. The numbers are still the numbers from 16 Nov 2026, 05:55 UTC, but they no longer describe the current production table. Read it as a dated observation, not a live rate.

What it proves

  • 86 rows in the Studio job table, grouped by the route and kind stored on the row.
  • 66 saved, 13 failed and 7 unknown, with every count adding back to the total.
  • Saved-job latency from created_at to the completed_at recorded on each saved row.
  • The cost estimate written on each row: US$12.86 across the window, before any provider charge.

What it deliberately does not

  • Rank quality. A saved job means an output was stored, not that it was good.
  • Rank quality across these routes. 3 recorded preference submissions on one matched pair is one unauthenticated comparison, not a preference across these 17 routes.
  • Call the estimate an invoice. No row in this window records a provider charge, so no invoice-derived cost is claimed.
  • Measure failure latency. Failed and unknown rows have no terminal timestamp, so they are counted, not timed.
  • Report token or credit usage. The generations table records no token or credit usage for these media jobs, so this page claims none.
  • Describe the provider, the account or the wider internet. It describes one Studio account in one short window.

Three inventories, one with job receipts

These stay separate on purpose. A build agent existing, a Worker binding being present and a media job finishing are three different observations. The job evidence below belongs only to the Studio media inventory.

build · agents

Build agents and harnesses

Which coding agents can a Composer user actually hand work to, at what published version, with which model ids the agent's own catalogue still lists?

Rows
13
Checked
16 Nov 2026, 14:34 UTC
State
Fresh · 5 days old

Current 11Drifted 1Unverified 1

Availability only. No job receipts are attached.

This inventory answers which agents and model ids exist. It has no completed job to report, so no saved rate is attached.

Open this inventory

hosted · execution

Cloudflare execution

What can the deployed App Composer Worker actually run today, which bindings does it really have, and which model ids, if any, is it able to execute?

Rows
12
Checked
16 Nov 2026, 02:41 UTC
State
Fresh · 5 days old

Current 8Not present 2Unverified 2

Availability only. No job receipts are attached.

This inventory answers what the deployed Worker can run and which bindings it has. A binding is not a job, so nothing is merged into the Studio rate.

Open this inventory

studio · media

Studio media endpoints

What can the independent Media Studio generate today, read from its own catalogue and then checked model by model against the provider's current published route?

Rows
21
Checked
16 Nov 2026, 02:41 UTC
State
Stale · 5 days old

Current 20Retired 1

Job receipts attached: 86 jobs on 17 routes.

The job receipts below are the Media Studio production table, mapped back to these endpoint rows. One production route is not listed in this inventory, and is labelled rather than invented.

Open this inventory

Saved rate, latency and recorded estimate by route

Sorted by job count. The saved rate is saved jobs divided by all jobs on that route. Median and p90 use saved jobs only, because they are the only rows with a completion timestamp. estimate recorded by Studio, not an invoice.

Media Studio jobs recorded 13 Nov 2026 to 15 Nov 2026. 16 routes map to a studio-media inventory row; 1 does not, and is labelled below.
Route and kindSavedSaved-job latencyEstimate median · total · not an invoice
MiniMax H3 Max Turbo · 768ph3-max-turboVideo100%23 of 23 savedno failures recorded19.4sp90 26.0s · 23 savedUS$0.24US$5.52 total
Hailuo 2.3 Standardhailuo-standardVideo76.5%13 of 17 saved4 unknown2m 01sp90 2m 39s · 13 savedUS$0.28US$4.76 total
Studio Assemblystudio-assemblyVideo22.2%2 of 9 saved4 failed · 3 unknown33.0sp90 41.5s · 2 savedUS$0.02US$0.18 total
Studio Detailstudio-detailImage44.4%4 of 9 saved5 failed7.7sp90 11.4s · 4 savedUS$0.006US$0.054 total
Studio Printstudio-printImage100%5 of 5 savedno failures recorded6.6sp90 11.6s · 5 savedUS$0.025US$0.125 total
Studio 9:16 Cutstudio-socialVideo25.0%1 of 4 saved3 failed28.7sp90 28.7s · 1 savedUS$0.03US$0.12 total
Studio Effectstudio-effectAudio100%3 of 3 savedno failures recorded14.3sp90 54.2s · 3 savedUS$0.008US$0.024 total
Studio Fillstudio-fillImage66.7%2 of 3 saved1 failed25.8sp90 26.0s · 2 savedUS$0.05US$0.15 total
Kling 2.5 Turbo Prokling-standardVideo100%2 of 2 savedno failures recorded2m 34sp90 2m 47s · 2 savedUS$0.35US$0.7 total
Studio Filmstudio-filmImage100%2 of 2 savedno failures recorded5.1sp90 5.5s · 2 savedUS$0.02US$0.04 total
Studio Scorestudio-scoreAudio100%2 of 2 savedno failures recorded1m 30sp90 2m 29s · 2 savedUS$0.012US$0.024 total
Studio Voicestudio-voiceAudio100%2 of 2 savedno failures recorded43.4sp90 1m 03s · 2 savedUS$0.06US$0.12 total
kling-o1-restylekling-o1-restyleVideoNot in the inventory100%1 of 1 savedno failures recorded5m 14sp90 5m 14s · 1 savedUS$1.01US$1.01 total
Studio Draft · FLUX.2 Klein 9Bstudio-craftImage100%1 of 1 savedno failures recorded4.1sp90 4.1s · 1 savedUS$0.016US$0.016 total
Studio Cutoutstudio-cutoutImage100%1 of 1 savedno failures recorded17.7sp90 17.7s · 1 savedUS$0.002US$0.002 total
Studio Draft · SDXL Lightning (retired)studio-draftImageRetired route100%1 of 1 savedno failures recorded5.0sp90 5.0s · 1 savedUS$0.0012US$0.0012 total
Studio Upscalestudio-upscaleImage100%1 of 1 savedno failures recorded1m 27sp90 1m 27s · 1 savedUS$0.02US$0.02 total
All 17 routes76.7%66 of 86 savednot pooledper-route medians stay per routeUS$12.86sum of recorded estimates

Latency is wall-clock time, including queue and provider time. It is not compute time and not time to first view. A route with one saved job has a median and p90 equal to that one job; treat it as a fact about that job, not a stable rate.

One route is missing from the inventory

kling-o1-restyle (Video) has 1 job in production, but no row in the 16 November studio-media inventory. The receipt keeps the production key and says so. It does not invent an endpoint or add an inventory row to make the tables agree.

The exact settings behind the numbers

The newest saved job of each kind in the same window. Only the settings listed on the row are shown; a setting the row did not record is omitted, not filled in.

Studio Effect

studio-effectAudio
Aspect
not recorded
Duration
5s
Pixels
not recorded

Studio Draft · FLUX.2 Klein 9B

studio-craftImage
Aspect
portrait
Duration
not recorded
Pixels
not recorded

Kling 2.5 Turbo Pro

kling-standardVideo
Aspect
from-image
Duration
5s
Pixels
not recorded

These are settings, not prompts or outputs. Pixel dimensions appear on exactly one job in the window: a failed Studio 9:16 Cut row at 720×1280. It contributes to the failure count, not to latency. Resolution is therefore not compared here.

How the receipt was read

  1. Read the schema first with PRAGMA table_info(generations). Every field named below exists on the production table; a field that did not exist was not inferred.
  2. Ran one read-only aggregate query per route and kind. It reads model, kind, status,created_at, completed_at, cancelled_at, cost_estimate_micro andprovider_charged_micro. It never selects request_json, a prompt, a project id, an asset id or a media key.
  3. Ran a second query for the three settings examples. It extracts only aspect, durationSeconds,aspectRatio, outputWidth and outputHeight from the request record. No other part of that record is read.
  4. Excluded the one row whose media was erased (deleted_at set). It is not counted and not described. No SELECT *appears in either query.
  5. Checked the committed aggregate with pnpm jobs:check. The page calls the same validator while building and refuses to render an undated file, a file with zero jobs, a mixed inventory, a private field, a negative latency or counts that do not add up.
Read the exact SQL

This is the aggregate query recorded in the evidence file. The examples query is recorded beside it.

WITH live AS (
  SELECT model, kind, status, created_at, cost_estimate_micro, provider_charged_micro,
         CASE
           WHEN status = 'saved' AND completed_at IS NOT NULL THEN completed_at
           WHEN status = 'cancelled' AND cancelled_at IS NOT NULL THEN cancelled_at
           ELSE NULL
         END AS terminal_at
  FROM generations
  WHERE deleted_at IS NULL
), saved AS (
  SELECT model, kind, terminal_at - created_at AS elapsed_ms,
         ROW_NUMBER() OVER (PARTITION BY model, kind ORDER BY terminal_at - created_at) AS rn,
         COUNT(*) OVER (PARTITION BY model, kind) AS n
  FROM live
  WHERE status = 'saved' AND terminal_at IS NOT NULL
), costs AS (
  SELECT model, kind, cost_estimate_micro,
         ROW_NUMBER() OVER (PARTITION BY model, kind ORDER BY cost_estimate_micro) AS rn,
         COUNT(*) OVER (PARTITION BY model, kind) AS n
  FROM live
)
SELECT l.model, l.kind,
  COUNT(*) AS jobs,
  SUM(l.status = 'saved') AS saved,
  SUM(l.status = 'failed') AS failed,
  SUM(l.status = 'unknown') AS unknown,
  SUM(l.status = 'cancelled') AS cancelled,
  MIN(l.created_at) AS first_created_ms,
  MAX(l.created_at) AS last_created_ms,
  SUM(l.terminal_at IS NULL) AS without_terminal,
  (SELECT (MAX(CASE WHEN s.rn = (s.n + 1) / 2 THEN s.elapsed_ms END) + MAX(CASE WHEN s.rn = (s.n + 2) / 2 THEN s.elapsed_ms END)) / 2.0
     FROM saved s WHERE s.model = l.model AND s.kind = l.kind) AS saved_median_ms,
  (SELECT MAX(CASE WHEN s.rn = (9 * s.n + 9) / 10 THEN s.elapsed_ms END)
     FROM saved s WHERE s.model = l.model AND s.kind = l.kind) AS saved_p90_ms,
  SUM(l.cost_estimate_micro) AS cost_estimate_sum_micro,
  (SELECT (MAX(CASE WHEN c.rn = (c.n + 1) / 2 THEN c.cost_estimate_micro END) + MAX(CASE WHEN c.rn = (c.n + 2) / 2 THEN c.cost_estimate_micro END)) / 2.0
     FROM costs c WHERE c.model = l.model AND c.kind = l.kind) AS cost_median_micro,
  SUM(l.provider_charged_micro IS NOT NULL) AS provider_charge_rows
FROM live l
GROUP BY l.model, l.kind
ORDER BY jobs DESC, l.model

Examples query:

WITH picked AS (
  SELECT kind, model,
         json_extract(request_json, '$.aspect') AS aspect,
         json_extract(request_json, '$.durationSeconds') AS duration_seconds,
         json_extract(request_json, '$.aspectRatio') AS aspect_ratio,
         json_extract(request_json, '$.outputWidth') AS output_width,
         json_extract(request_json, '$.outputHeight') AS output_height,
         ROW_NUMBER() OVER (PARTITION BY kind ORDER BY created_at DESC) AS rn
  FROM generations
  WHERE deleted_at IS NULL AND status = 'saved'
)
SELECT kind, model, aspect, duration_seconds, aspect_ratio, output_width, output_height
FROM picked
WHERE rn = 1
ORDER BY kind

Where this comparison stops