Kaelvorn: build and redesign benchmark

One Claude Code orchestrator on Claude Opus 5.5 commissioned research, art, implementation and reviews through the delegate skill. The result is the live Kaelvorn site: 37 prerendered pages for a fictional upcoming MMORPG, with 13 Callings, four full dossiers, skill trees and a painted atlas. The redesign brought the implementation closer to its painted concepts, adding a central HUD, a heraldic logo, brush masks, crown moulding and painted interface icons.

Summary

These are Standard list-price usage valuations, not subscription invoices. Image output uses the first benchmark's high-quality assumption and explicitly declared size proxies; reasoning is not charged twice. All values are rounded only for display. Agent time sums overlapping run durations, not elapsed delivery time; orchestrator hours are unmeasured.

The fixed source snapshot is usage.json, orchestrator.json, prices.md (sources read 2026-10-26), the decision log, image ledger, redesign plan and benchmark brief. Build means the recorded codex-1…25, antigravity-1…9, cursor-1…6 and gemini-api-1…3 records, including cursor-2-nested-codex. There is no cursor-3 record. Redesign means codex-26…33, cursor-7…12 and antigravity-10…13. This writer (codex-34) is absent and excluded; 61 runs are present, including four quota attempts.

Cost and usage per agent

Each stage cell reports USD; runs / turns; time (hours); fresh / cached / output / reasoning tokens (F/C/O/R); images. Image-model rows split output charges from their owners' text costs without adding delegations or hours: Nano Banana was invoked inside antigravity-7 in the build and cursor-7…10 in the redesign; Codex image calls were inside Codex runs. Thus text rows show zero images allocated to that row, rather than claiming the harness generated none.

¹ Text formula and tier. For each run, USD = (F × fresh rate + C × cache-read rate + O × output rate) / 1,000,000. As in the first report, input is treated as fresh input and cached as separate cache reads; cached tokens often exceed input, so subtraction would be invalid. Rates F/C/O are GPT-6.1 Sol 2 / 0.10 / 10, Cursor GPT-5.6 Sol 4 / 0.40 / 20, Cursor Grok 4.7 2 / 0.50 / 6, and Gemini 4 Argon 2 / 0.10 / 10 (the promotional Standard tariff valid through 2026-11-30), from prices.md. Model effort suffixes are normalized to the same model; no Fast multiplier is assumed. Output is charged once; reported reasoning is treated as included in output. Zero reasoning means none reported.

Long-context estimate, following the first benchmark's convention: use fresh input divided by recorded turns as a proxy for per-turn input. The whole-run GPT-6.1 Sol long tier 4 / 0.20 / 15 applies to cursor-2-nested-codex (595,275), codex-30 (1,380,020 / 5 = 276,004), and codex-33 (275,141), above 272,000. Actual tiers depend on individual API requests including cached context; aggregate run data cannot establish them. This preserves the original estimator, not proof of billing. No separately reported agent cache-write counts are available.

² Nano Banana formula. Standard 2K image output = images × 1,680 image-output tokens × $30 / 1,000,000 = images × $0.0504. Build: 8 × 0.0504 = $0.4032000; redesign: 20 × 0.0504 = $1.0080000; total $1.4112000. Prompt/reference inputs and image-model text/thinking output are unreported and excluded; image calls are not extra delegation runs.

³ Codex image formula. GPT Image 2 high-quality output-only valuation: 1024×1536 and 1536×1024 $0.165; saved 1672×941/940 uses nearby valid 1680×944 $0.13944; saved 1254×1254 uses 1248×1248 $0.24591. The last two are calculator proxies, not exact tariffs for those files. Actual generation quality and image input tokens are unavailable. Build: 55 × 0.13944 + 46 × 0.165 + 2 × 0.24591 = $15.7510200; redesign: 6 × 0.13944 + 6 × 0.165 + 4 × 0.24591 = $2.8102800; total $18.5613000.

⁴ Orchestrator formula and boundary. Use the explicitly supplied by_stage prices: build $24.20, redesign $37.42, sum $61.62. Rates F/cache-read/1h-cache-write/output are 4 / 0.20 / 8 / 20 USD/M; 5-minute cache writes are zero. The unrounded recomputations are $24.2031742 and $37.4160834, consistent with the rounded stage prices. The older top-level snapshot says $61.41 and 455 responses; the stage records sum to 456 responses and have newer token totals. We use the requested stage split consistently, not that stale top-level total. This is one session across both stages, not two sessions; responses are not delegate turns. The redesign includes writing and fixing the delegate skill's plan workflow in that same session, so it is not solely site-production cost. The JSON links its Claude price source.

Shares and rounding: orchestrator share = stage orchestrator USD / (unrounded agent USD + stage orchestrator USD) × 100. Currency totals sum unrounded text and image costs, so rounded rows may differ by a cent from displayed totals.

Recomputable text subtotals

Each line below sums exactly the matching run IDs in usage.json under the formula above. F/C/O/R counts and hours are in the cost table; the explicit IDs and rate selections remove ambiguity about grouping.

The long-context uplift is $3.2097834 in the build and $17.2533492 in the redesign. An all-base GPT-6.1 Sol sensitivity reduces both-stage agent cost by $20.46 to $120.49, and combined cost to $182.11; this changes no usage counts. The original build report was $72.05 agents / $23.24 orchestrator: the current build includes codex-25's second turn (639 seconds versus 501, with refreshed tokens) and the updated $24.20 stage snapshot. Those snapshot changes explain the revised build figures.

Images: attempts, not files

Generated outputs include rejected attempts. Cropping icons, keying backgrounds, favicon exports, contact sheets, atlas pin previews, deterministic SVGs and delivery derivatives are not additional generations. The image ledger records the attempts; the decision log confirms the Nano Banana reply counts.

Navigation has six first passes and five remakes; one rejected mobile first pass was actually landscape 1672×940. Logo concept C needed a second attempt to retain the full KAELVORN spelling. The atlas was retained on its first generation. Nano Banana redesign counts reconcile to 3 + 7 + 6 + 4 = 20: icons discarded one sheet; masks' rejected 5:1 requests generated nothing; ornaments had no remakes; the horizontal logo needed one remake. Icon validation also rejected one request before generation. These failed requests add no images.

The build's 103 Codex generations yielded 63 retained originals (seven style references and 56 production paintings); eight Nano Banana generations yielded six sheets, cut into 48 skill and seven item icons. The original filesystem reconciliation remains 103 − 40 + 8 − 2 + 55 = 124 build PNGs, including six sheets and seven style references, versus 111 production assets. That historical count is not the redesigned repository's current PNG count.

Who did what in the redesign

The original plan had 11 units; atlas became a new unit and benchmark followed publication. Times below are complete recorded run durations, including follow-ups, from usage.json; outcomes come from DECISIONS.md, starting at “Redesign (phase 10)”, and the plan. Zero recorded duration on quota attempts does not establish zero actual waiting time.

The benchmark unit itself is this report; no codex-34 usage record exists in the supplied snapshot. Redesign Antigravity records contain four attempts, zero recorded turns/seconds and no token fields: $0.00 recorded is a lower-bound accounting entry, not evidence that the failed harness sessions consumed no resources.

What the plan workflow changed

Checked consumers before launch. plan check caught two outputs with no consumer before wave 1 started. The decision log does not name the two files, so they cannot be identified from these inputs; the useful change was requiring an explicit handoff for every output, instead of assuming generated work would be integrated.

Reassignment kept the work moving. Antigravity exhausted its account quota, with a reported reset roughly 125 hours away; audit, ornaments, masks and ui-icons could not finish there. The plan moved audit/review to Codex with headless Chrome and image-making units to Cursor GPT-5.6 Sol, retaining Nano Banana through gemini-api image; logo-final changed assignment before an Antigravity run was launched.

Send-backs preserved context and respected new code. The first visual review's 19 findings went back to codex-30 for shared components, while page fixes started on a fresh Cursor worktree because cursor-11 predated the shared changes. Re-review still found three regressions—mobile launch art below the fold, fragmented “Revealed soon” text and copy over brush fringes—so cursor-12 received another follow-up; the orchestrator also returned clipped/doubled header branding to its shell owner.

Left open became an owned deliverable. The SVG unit's README exposed that the audit called for a painted world atlas, but no unit made it. The orchestrator added atlas, accepted its 3:2 painting and six pin positions, then sent it to the existing shell builder as an attached follow-up; acceptance by consumers also determined when upstream units closed.

Incremental application needed an explicit contract. The benchmark brief identifies incremental diff --apply as a skill fix discovered during the workflow: repeated follow-ups require applying the new delta against already accepted work. The supplied decision log does not describe its implementation or verification, so this report records the fix without claiming a particular patch algorithm or test result.

Follow-up timeouts must inherit the run's settings. codex-30's third turn hit 30 minutes because send ignored the timeout used by start; the decision log records a skill fix. A follow-up is part of the same long-running unit and needs its execution settings carried forward along with its context.

Completion needs a parseable status line. The benchmark brief lists a status-line skill fix, and the delegate brief requires a final STATUS: done, STATUS: question … or STATUS: blocked … line. That gives the orchestrator an explicit terminal outcome beside the deliverables and Left open list; the supplied sources do not detail the parser change.

Nobody approves plans mid-turn. cursor-12 initially ended empty after creating a plan for approval, so it was sent back; the skill now tells delegates nobody approves plans during their turn. Delegates must execute within the brief, or finish with a concrete question when an orchestrator decision is essential, rather than wait for an interactive approval that cannot arrive.

Who did what in the build

The first stage's complete run table is retained below. Durations now use the refreshed usage snapshot (notably codex-25's two turns); the original per-phase acceptance and rework notes remain useful context. Units often ran in parallel.

Build lessons retained

Cursor delegated the builder’s entire job. Builder C read the orchestrator’s CLAUDE.md, assumed that role and launched a 1,475-second Codex worker; its implementation was accepted, but its decision-log edits were discarded. The delegate lesson is to make role and scope explicit in every brief: subsequent briefs say the agent is a delegate, must do the work itself and must not edit the decision log; account for nested workers under their actual harness.

A foreground server consumed the browser reviewer’s turn. The decision log says Antigravity’s first review turn timed out while running vite preview in the foreground, and a second turn wrote findings from 46 screenshots. The skill lesson is to start persistent servers in the background and preserve evidence between turns; the snapshot records only one turn for antigravity-8, so it cannot quantify the missing attempt separately.

Builder B found a production hydration deadlock in A’s shell. A top-level await import() in preferences.tsx blocked production hydration even though earlier checks had passed. The fix became a separate Codex run with npm run smoke: delegated handoffs need a built-site browser check of real interaction, not just compilation and unit tests.

Budget gates stopped integration twice. Builder A’s shell measured 102.4 KiB gzip against a 100 KiB gate; the orchestrator raised it to 110, then later raised home/static route gates from 160/150 to 175/160 after measuring 166.5/150.6. The lesson is to give delegates a measured dependency baseline and an escalation path for budget changes, then require performance evidence: Lighthouse subsequently identified a real late-CSS defect rather than treating the relaxed gates as proof of acceptable UX.

Ownership boundaries still needed an integration pass. B’s WorldMap brought its own h2 into C’s /world section, breaking one integration test; the integration worker corrected the page structure instead of changing the assertion. The skill should specify shared component contracts, merge checkpoints and an integration owner; phase 8 then split copy, A/B and C fixes by ownership so parallel agents would not edit the same files.

Reviews and fixes needed more than one round. Three phase 7 reviewers returned 17 marketing, 12 browser and 9 performance findings; fixes reached 257 tests and 11 smoke checks and raised character-page desktop Lighthouse from 79 to 100, but the browser/Grok re-checks found incomplete fixes and new regressions. After a final fix run, the orchestrator’s own visual pass still found four polish issues, so the lesson is to budget for independent re-checks and a final visual acceptance pass, with concrete routes, widths and evidence in every fix brief.

Art delivery hid substantial iteration. All 56 phase 5 Codex finals were accepted by the orchestrator, yet the complete ledger records 40 discarded Codex attempts across style and production work, plus two discarded Nano Banana sheets. The skill should preserve per-attempt image counts, native dimensions and rejection reasons; final file counts alone understate generation cost and the effort needed for consistent art.

Scope, timing and verification limits

• Elapsed versus summed time: the original build's commit bracket ran from 2026-10-25 23:07:26 to 2026-10-26 03:49:19 (+02:00), 4h 41m 53s; including its first benchmark turn estimated roughly 4h 51m. Those are historical estimates, not updated stage wall times. usage.json has durations without start/end timestamps, so redesign elapsed time and active orchestrator hours cannot be recovered.
• Nested overlap: cursor-2's 1,599 seconds contain its 1,475-second nested Codex worker. Subtracting that interval once yields build 8h 43m 02s, and both stages 13h 52m 14s; this still is not wall time because other workers overlapped. Runs count unique delegation records, turns include recorded follow-ups.
• Snapshot gaps: the decision log describes more cursor-12 follow-up activity than its two recorded turns, and Antigravity's build browser retry is not fully represented by its turn count. Outcomes follow the final decision log; quantitative usage follows usage.json without inventing missing tokens/time. The current writer and later orchestrator work are outside the snapshot.
• Valuation limits: token categories vary by harness; Standard speed and the stated approximate context tiers are assumptions. Excluded items include separately unreported cache writes, image prompt/reference and image-model text/thinking charges, tool/search charges, taxes and subscription fees. Agent cost_usd zeros in the JSON are not invoices.
• Evidence of the result: the decision log records publication and every redesign unit closed. Re-review was not a claim that every visual issue disappeared: F13/F18 were partly accepted, and the last fixes were accepted by the orchestrator; this report does not claim a further independent post-publication audit.