BuilderBenchby the LAB
8 October 2026 · v3 · pack 3.0.0
4-hour build window · Extra High (xhigh) · Standard
the reveal · owner-finalized partial personal assessment
One owner-finalized partial personal assessment of one GPT-6 Luna run with Codex and one Haiku 5.5 run with Claude on the v3 pack. This is not a universal model ranking or a statistical claim.
Codex harnessleads this assessment
44.58/100
API estimate in US dollars · partial usage · 18 of 18 records priced
API equivalent · not a bill
Claude harnessfinal score
39.18/100
API estimate in US dollars · partial usage · 20 of 21 records priced
API equivalent · not a bill
owner-finalized partial personal assessment. one run per setup. a lead in one assessment is not a ranking. the limits are below.
quality is the weighted average of the task scores. autonomy is how little owner rescue the setup needed. proactivity is useful initiative inside the brief. each is scored out of 100.
the final score is quality × 70% + autonomy × 20% + proactivity × 10%. cost never buys points.
dated references
each bar is one owner review of one run. the packs, tasks and evidence limits differ, and each bar says which pack and which kind of review it was. scores only compare inside one pack.
task breakdown
each task is scored out of 100. its weight is its share of the quality score.
| task | GPT-6 Lunascore /100 | Haiku 5.5score /100 |
|---|---|---|
| 01 · Writing10% of quality | 18 | 21 |
| 02 · Landing page10% of quality | 30 | 30 |
| 03 · App lead magnet20% of quality | 50 | 48.75 |
| 04 · Video editing10% of quality | 11.67 | 5 |
| 05 · Motion5% of quality | 15 | 10 |
| 06 · Thumbnails10% of quality | 12 | 26 |
| 07 · Browser + computer10% of quality | 55 | 55 |
| 08 · Research10% of quality | 84.50 | 82 |
| 09 · Automation10% of quality | 100 | 83.33 |
| 10 · Founder conversations5% of quality | 65 | 52.50 |
| task | GPT-6 LunaAPI-equivalent estimate | Haiku 5.5API-equivalent estimate |
|---|---|---|
| Writing | US$0.03recorded usage | US$1.31recorded usage |
| Landing page | US$0.14recorded usage | US$6.11recorded usage |
| App lead magnet | US$0.22recorded usage | US$5.04partial usage |
| Video editing | US$0.48recorded usage | US$9.11recorded usage |
| Motion | US$0.42recorded usage | US$7.24recorded usage |
| Thumbnails | US$0.10recorded usage | US$4.58recorded usage |
| Browser + computer | unavailableusage unavailable | unavailableusage unavailable |
| Research | US$0.03recorded usage | US$1.42recorded usage |
| Automation | US$0.05recorded usage | US$2.59recorded usage |
| Founder conversations | unavailableusage unavailable | unavailableusage unavailable |
estimates, not bills. “unavailable” means no usage was recorded for that task. it does not mean the task was free.
native sessions
these four sessions run in the real app, outside the build window, and each gets its own rating. browser and computer average into the browser + computer task. advice and planning average into the founder conversations task. see the four prompts
API pricing
the rates are what the provider listed on the date shown. the estimate prices the usage this run recorded. the bill is what was actually charged. they are kept apart on purpose.
| per 1M tokens | input | cache read | output |
|---|---|---|---|
| Up to 272K input | $0.10 | $0.01 | $0.50 |
| Above 272K input | $0.20 | $0.02 | $0.75 |
Cache writes: $0.125 / $0.25. Tool and regional fees are separate. The run estimate uses the short-context row; request-level bands and cache writes were unavailable.
US dollars · Standard processing · OpenAI API docs: GPT-6 Luna · checked
| per 1M tokens | input | cache read | output |
|---|---|---|---|
| Up to 100K input | $0.10 | $0.01 | $0.50 |
| Above 100K input | $0.50 | $0.05 | $2.50 |
5-minute cache writes: $0.125 / $0.625. 1-hour cache writes: $0.20 / $1.00. Tool fees are separate. The run estimate keeps provider-reported figures instead of applying these rates.
US dollars · Standard processing · Claude API docs: Haiku 5.5 · checked
API-equivalent estimates price recorded execution usage only. They are not subscription bills. GPT-6 Luna uses a dated Standard short-context token scenario. Haiku 5.5 keeps provider-reported estimates, including recorded tools. Both are partial and use different methods, so they do not support an exact cost ratio. Native sessions, unknown cache writes, context premiums, regional fees, setup, calibration, coordination, grading, presentation and subscription fees are excluded. Actual incremental spend is unknown.
2026-10-11 05:30 UTC · fixed after this result went public
what happened: the result graphic printed "candidate A" and "candidate B" under the model names, and a stray dash after "extra high".
why: leftovers from the blind stage of the image. the two labels and the dash were taken out. the first graphic stays online at /builderbench/2026-10-08-luna-haiku/luna-haiku-reveal.png.
what did not change: every score, margin, price, date and caveat. no number in the graphic or on this page changed.
review conditions
Requested as a blind comparison. The build work was rated as Candidate A and Candidate B, and the names were revealed after score lock. It was not perfectly blind: earlier operator diagnostic output included cost clues. The four native sessions were rated by model name in a separate form that never shows which candidate is which.
The owner rated the saved build outputs, the four native sessions, autonomy and proactivity. Task 9 technical repair is conductor-scored with recorded evidence. For this run only, the owner also delegated grading of the research memo and decision table to the conductor. The lead-magnet app keeps this run's original frozen four-part rubric. The saved scores were finalized as a partial personal assessment and locked. Quality is the weighted task score. v1 to v3 use 70% quality, 20% autonomy and 10% proactivity.
frozen setup
exact model and harness version strings are not in this public export, so they stay unknown here.
limits
p.s.
live count, 16 oct 2026 / 17 oct 2026.