Benchmarks Studio · inventories

What is true today, and when we last looked.

Three separate questions, three separate answers. 46 rows, each naming the exact source it was read from, the status that source supports and the date of the read. They are data: a fetched page describes a product, it never becomes permission or a command.

Rows checked
46
Inventories
3
Unverified rows
3
Stale inventories
1
Paid trials run by this page
none

How to read a row

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?

FreshChecked 2026-11-16T14:34:47Z5 days ago · stale after 7 days
Current
11
Drifted
1
Unverified
1
Changed since last check
3

What changed since the previous check

Previous check 2026-11-15T08:42:03.193Z from benchmarks/results/compare.json. The comparison run recorded the three configurations it executed: Codex CLI 0.145.0 with gpt-5.6-sol and gpt-5.6-terra, and an ingested OpenCode agent trial.

  • codex-cli · versionwas 0.145.0 (the harness the comparison run recorded on 2026-11-15)now 0.169.0 published 2026-11-13T08:37:05Z; the local install is still 0.145.0The measured harness is 24 published minors behind. A new timed comparison has to name the version it runs, or the result is not comparable with the recorded one.
  • codex-gpt-6-1-sol · capabilitywas gpt-5.6-sol and gpt-5.6-terra were the only Codex models the track had recordednow eight listed Codex ids, with gpt-6.1-sol as the operator's selected defaultThe two measured models are still offered, so the recorded comparison is not invalidated; it is no longer the newest pair.
  • opencode · presencewas Earlier comparison recorded a host agent with no CLI version.now Desktop 2.2.7 bundled CLI and Homebrew CLI 2.3.0 were both observed.Short inference checks on the bundled CLI used both HY4 Flash Free routes.

Sources for this check

  • https://github.com/openai/codex/releasesGitHub releases API 200: tag_name rust-v0.169.0, published_at 2026-11-13T08:37:05Z · read 2026-11-16T02:24:00Z
  • file:~/.codex/models_cache.jsonThe client's own cache of the vendor model list: fetched_at 2026-11-16T02:22:50.619Z, client_version 0.164.0, etag present, ten records with slug/visibility/supported_in_api · read 2026-11-16T02:22:50Z
  • https://developers.openai.com/codex/models200, 634,565 bytes; the page names gpt-6.1-sol, gpt-6-astra and gpt-6.1-terra · read 2026-11-16T02:34:00Z
  • https://developers.openai.com/codex/changelog200; the changelog names gpt-6.1-sol, gpt-5.6-sol and gpt-5.6-terra · read 2026-11-16T02:34:00Z
  • https://github.com/sst/opencode/releasesGitHub releases API 200: tag_name v1.19.5, published_at 2026-10-28T06:17:49Z · read 2026-11-16T02:26:00Z
  • https://registry.npmjs.org/@anthropic-ai%2Fclaude-code/latestnpm registry 200: version 2.1.325 · read 2026-11-16T02:26:00Z
  • https://models.dev/api.json200, 6,654,265 bytes: 324 providers; opencode-go has 40 models including deepseek-v4.1-flash · read 2026-11-16T02:26:00Z

What this inventory still cannot prove

  • The initial inventories were checked on 16 November 2026. Each row retains its own check date, including the later OpenCode observation.
  • HY4 Flash Free was exercised with short inference prompts. Benchmark results remain attached to their original run and configuration.
  • Two further Codex catalogue records are marked visibility "hide" (gpt-reserve, codex-auto-review) and are not asserted as usable here.
  • The Codex catalogue was written by client 0.164.0, while the installed CLI reports 0.145.0. The list is the vendor's current published catalogue for the account, not a capability claim about the older binary.
  • Five of the eight listed Codex ids are also named on the vendor's public models page or changelog. gpt-6-sol, gpt-6-luna and gpt-5.6-luna appear only in the client's freshly fetched catalogue and are recorded on that evidence alone.
  • Claude Code remains unverified on the observed host.
  • OpenCode lists HY4 Flash Free as a limited-time offer.

Pinned runs, left as recorded

  • compare-v12026-11-15T08:42:03.193Z · file:benchmarks/results/compare.jsonNine tries of one frozen brief across three configurations. Left exactly as recorded; this inventory only describes the harnesses around it.

hosted · execution

Cloudflare-hosted Composer 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?

FreshChecked 2026-11-16T02:41:00Z5 days ago · stale after 7 days
Current
8
Not present
2
Unverified
2
Changed since last check
0

What changed since the previous check

No earlier recorded check exists for this inventory, so there is nothing to diff yet. That is a fact about it, not an omission.

No difference was found in the rows this inventory compares.

Sources for this check

  • https://composer.loftwah.com/build.json200 JSON: app, commit, reviewed, sourceCommittedAt, contract · read 2026-11-16T02:20:00Z
  • https://api.cloudflare.com/client/v4/accounts/1003dc1d93af0ebea56b2f1252f89627/workers/scripts/app-composer/settings200 JSON: result.bindings[] with type and name; result.compatibility_date 2026-10-13; compatibility_flags nodejs_compat, global_fetch_strictly_public · read 2026-11-16T02:27:00Z
  • https://api.cloudflare.com/client/v4/accounts/1003dc1d93af0ebea56b2f1252f89627/workers/scripts/app-composer/deployments200 JSON: ten deployments, each with id, created_on, source "wrangler", versions[].version_id · read 2026-11-16T02:27:00Z
  • https://api.cloudflare.com/client/v4/accounts/1003dc1d93af0ebea56b2f1252f89627/ai/models/search403 JSON: success false, errors[0].code 10000, message "Authentication error" · read 2026-11-16T02:28:00Z

What this inventory still cannot prove

  • This is the first recorded check for this plane, so there is no previous value to diff. The changes list is empty by fact, not by omission.
  • The inventory describes deployed commit 8d5d0c5. Three commits on origin/main are not deployed, so a later release invalidates the deployment row and nothing else here without a fresh read.
  • Secret values were never read. Only the binding names and types from the Worker's own settings response are recorded.
  • Whether a workspace has an enabled media connection is customer data and was not queried.
  • Account-level Workers AI eligibility is unverified because the available token is refused by that endpoint; no eligibility claim is made for any model id.

studio · media

Studio media models

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?

StaleChecked 2026-11-16T02:41:00Z5 days ago · stale after 3 days
Current
20
Retired
1
Changed since last check
3

What changed since the previous check

Previous check 2026-11-15T12:00:22.518Z from config/media-studio-capabilities.json (Composer's committed copy of https://media.loftwah.com/capabilities.json). Composer's snapshot of the Studio manifest, refreshed after the Studio Draft retirement. It listed the eight Composer-selectable models and one retired record.

  • h3-max-turbo · pricewas a 30% promotional rate of $0.028/s at 768p ran through 2026-11-15, so the 5 s allocation of $0.24 carried a promotion caveatnow from 2026-11-16 the standard rate applies: $0.04/s at 768p. The allocation already used that standard rate, so recorded takes do not need re-pricingRead from the provider's own price description today. It is the one real change since the previous check, and it changes a caveat rather than a number.
  • studio-detail · capabilitywas Composer's committed snapshot recorded 8 models, because only selectable ones are storednow the Studio's own catalogue offers 16 models today, of which 8 are selectable in ComposerThe eight extra routes are real and runnable in the Studio. A reader who only saw Composer's snapshot would understate what the Studio can do.
  • studio-draft-retired · availabilitywas retirement published for 2026-11-20, replacement already shipping (recorded 2026-11-15)now 2026-11-16: still resolvable, four days remain, replacement studio-craft still currentThe window is closing. If the page starts returning 404 before 20 November, the retired record stays and only the observation changes.

Sources for this check

  • https://media.loftwah.com/capabilities.json200 JSON: contract, version, checkedAt, lastGoodAt, models[16], retired[1], frames, adapters, sources[18], alerts[] · read 2026-11-16T02:18:00Z
  • https://github.com/loftwah/composer-media-studio/blob/f2f29696d8be2424cee5f64b89bdcb0bb43de037/src/lib/providers.mjspublicModels() 16 records and publicRetiredModels() 1 record; file SHA-256 2d45ea17…6f4 at the deployed revision · read 2026-11-16T02:22:00Z
  • https://developers.cloudflare.com/workers-ai/models/Five live ids and the retired one fetched as /models/<slug>/index.md: all HTTP 200; only stable-diffusion-xl-lightning carries Planned Deprecation 11/20/2026 · read 2026-11-16T02:22:00Z
  • https://fal.ai/api/modelsEleven keyword searches, one per fal endpoint: every one returned the exact id with status "public", deprecated false, removed false · read 2026-11-16T02:23:00Z
  • file:config/media-studio-capabilities.jsonEight selectable models, one retired record, checkedAt 2026-11-15T12:00:22.518Z; identical to the live manifest on endpoint, kind, costMicro and retired record · read 2026-11-16T02:24:00Z

What this inventory still cannot prove

  • No generation ran for this inventory. Every row is a catalogue read, so a live route is not proof that a request would succeed, cost what the estimate says, or look right.
  • Planning costs are the Studio's own estimates. No provider invoice was read, and none is claimed.
  • The Studio manifest's own checkedAt is 2026-11-15T12:00:22.518Z. The per-model verification here is a separate read on 2026-11-16, so the two dates differ by design.
  • Only one Workers AI route in this inventory has an on-account generation receipt (studio-craft). The other Workers AI rows are catalogue-verified: the provider page resolves, which is not the same as a successful request on this account.
  • fal's published price description for h3-max-turbo is time-bounded prose. It changed meaning on 2026-11-16 and will change again.
  • A withdrawn id can disappear from fal search without a deprecation flag. Every endpoint is matched by exact id, and a missing exact match is recorded as drift rather than assumed.

Pinned runs, left as recorded

  • kling-vs-h3-202611152026-11-15T10:49:26Z · url:https://github.com/loftwah/app-composer/issues/601#issuecomment-6893211413One matched still-to-motion pair through the normal Create UI: H3 gen_a403ff, allocation $0.24, 1344x768, 5.18 s; Kling gen_1e7d9f9e, allocation $0.35, 1920x1080, 5.04 s. Provider charge not reported for either. Left exactly as recorded.
  • studio-craft-generation-202611152026-11-15T12:00:22.518Z · file:docs/evidence/launch-20261115/model-freshness-20261115.mdThe recorded on-account generation for studio-craft: gen_4515ab469ca948dfbe5587bbf142c8ac, 1024x1280 JPEG, 459,080 bytes, SHA-256 b32781d2…. It is the only Workers AI route in this inventory with an on-account receipt.
  • composer-media-snapshot2026-11-15T12:00:22.518Z · file:config/media-studio-capabilities.jsonThe snapshot this check compared against. Not rewritten by this change.

Queued, not run

What this inventory cannot answer without a decision

Each row names its cost label and the action only the owner can take. Nothing on this page has been submitted, generated or billed.

4 queuedChecked 2026-11-16T02:41:00Z

Where these inventories stop