Lennox Saint
Open menu

BuilderBenchby the LAB

10 November 2026 · v3 · pack 3.1.0
4-hour build window · Extra High (xhigh) · Standard

full result · owner-finalized partial personal assessment

Opus 5.5 scored 74.25. Fable 5.2 scored 69.43.

One owner-finalized partial personal assessment of one Fable 5.2 run and one Opus 5.5 run in Claude (Claude Code for the build, the Claude desktop app for the live sessions) on pack 3.1.0. This is not a universal model ranking or a statistical claim.

words used on this page, in plain English
test (pack)
a fixed set of tasks. every AI in the same pack gets the same tasks, so scores only compare inside one pack.
score out of 100
70% how good the finished work was (quality), 20% how little help it needed (autonomy), 10% useful extras nobody asked for (proactivity).
harness
the app or tool that runs the AI, like Codex or Claude Code. the score belongs to the AI and its harness together.
cleared result
a result Lennox checked and approved for publishing.
owner review
Lennox reviewed and scored the work. some checks are automated, and on one result he asked Claude to rate some parts for him. each result page says which. each result is one review of one run per AI.
named or blind review
named: Lennox could see which AI made the work. blind: the names were hidden until the scores were saved.
partial review
some of the work could not be fully checked. the result page says which parts.
API estimate
what the run would have cost at the AI company's listed prices. it is an estimate, not what was paid.
reasoning (xhigh)
how hard the AI was set to think. xhigh means extra high, the highest setting used.

final scores out of 100

full result · BuilderBench test of 10 November 2026 · 2 AIs on this page

Claude harnesswon this test

Opus 5.5

74.25/100

quality 70% of the score
73.79
autonomy 20%
73
proactivity 10%
80
US$88.82

API estimate in US dollars · partial usage · 18 of 18 records priced
API equivalent · not a bill

full result · BuilderBench test of 10 November 2026 · 2 AIs on this page

Claude harnessfinal score

Fable 5.2

69.43/100

quality 70% of the score
65.34
autonomy 20%
77
proactivity 10%
83
US$38.71

API estimate in US dollars · partial usage · 12 of 18 records priced
API equivalent · not a bill

Opus 5.5 leads by 4.82 points.

owner-finalized partial personal assessment. one run per setup. winning one test does not make an AI the best overall. the limits are below.

how each score is built

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.

  • Opus 5.5: 51.65 + 14.60 + 8.00 = 74.25
  • Fable 5.2: 45.74 + 15.40 + 8.30 = 69.43

dated references

every cleared score, by date. not a leaderboard.

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.

  1. Opus 5.510 Nov 2026 · v3 · pack 3.1.0 · with Claude
    owner-finalized partial personal assessment · this result
    74.25
  2. Fable 5.210 Nov 2026 · v3 · pack 3.1.0 · with Claude
    owner-finalized partial personal assessment · this result
    69.43
  3. Gemini 4 Neon19 Nov 2026 · v3 · pack 3.1.0 · with BuilderBench API harness
    owner-finalized partial personal review, with 3 of 28 ratings made by the bench's conductor (claude opus 5.5) and disclosed
    58.76
  4. Gemini 4 Hydrogen12 Nov 2026 · v3 · pack 3.1.0 · with BuilderBench API harness
    owner review, with 7 of 28 ratings made by claude and disclosed
    51.11
  5. Gemini 4 Carbon11 Nov 2026 · v3 · pack 3.1.0 · with BuilderBench API harness
    owner-finalized partial personal assessment
    49.04
  6. GPT-6.1 Terra27 Oct 2026 · v3 · pack 3.0.0 · with Codex
    owner-finalized partial personal review
    48.88
  7. Kimi K3.1 Pro27 Oct 2026 · v3 · pack 3.0.0 · with BuilderBench API harness (OpenRouter)
    owner-finalized partial personal review
    46.68
  8. Gemini 4 Argon14 Oct 2026 · v3 · pack 3.0.0 · with BuilderBench API harness
    owner-finalized partial personal review
    45.38
  9. Muse Wonder Full14 Oct 2026 · v3 · pack 3.0.0 · with Muse Code
    owner-finalized partial personal review
    39.91
  10. GPT-6 Luna8 Oct 2026 · v3 · pack 3.0.0 · with Codex
    owner-finalized partial personal assessment
    44.58
  11. Haiku 5.58 Oct 2026 · v3 · pack 3.0.0 · with Claude
    owner-finalized partial personal assessment
    39.18
  12. GPT-6.1 Sol30 Sep 2026 · v3 · pack 3.0.0 · with Codex
    owner-finalized partial personal assessment
    41.75
  13. Sonnet 5.529 Sep 2026 · v3 · pack 3.0.0 · with Claude
    owner-finalized partial personal assessment
    48.98
  14. GPT-6 Astra26 Sep 2026 · v2 · pack 2.0.0 · with Codex
    personal assessment with accepted recording limitations
    53.75
  15. Opus 5.524 Sep 2026 · v1 · pack 1.0.0 · with Claude
    one owner-scored pilot
    68.26
  16. GPT-6 Sol24 Sep 2026 · v1 · pack 1.0.0 · with Codex
    one owner-scored pilot
    45.55

task breakdown

where each setup gained and lost ground

each task is scored out of 100. its weight is its share of the quality score.

swipe the table sideways to see every column.

Task scores out of 100 for each setup, with each task’s share of the quality score
taskOpus 5.5score /100Fable 5.2score /100
01 · Writing10% of quality51.4039.80
02 · Landing page10% of quality7040
03 · App lead magnet20% of quality7157
04 · Video editing10% of quality7876.67
05 · Motion5% of quality8268
06 · Thumbnails10% of quality77.2076.80
07 · Browser + computer10% of quality7275
08 · Research10% of quality8071
09 · Automation10% of quality8989.33
10 · Founder conversations5% of quality74.5073.50
see the API estimate for each task

swipe the table sideways to see every column.

API-equivalent estimate and usage coverage for each task and setup
taskOpus 5.5API-equivalent estimateFable 5.2API-equivalent estimate
WritingUS$3.44recorded usageUS$2.49recorded usage
Landing pageUS$15.24recorded usageUS$4.81partial usage
App lead magnetUS$26.95recorded usageUS$11.36partial usage
Video editingUS$17.07recorded usageUS$7.64partial usage
MotionUS$8.57recorded usageUS$5.87recorded usage
ThumbnailsUS$5.23recorded usageUS$2.16partial usage
Browser + computerunavailableusage unavailableunavailableusage unavailable
ResearchUS$2.84recorded usageUS$1.15partial usage
AutomationUS$9.48recorded usageUS$3.23partial usage
Founder conversationsunavailableusage unavailableunavailableusage unavailable

estimates, not bills. “unavailable” means no usage was recorded for that task. it does not mean the task was free.

native sessions

browser, computer, advice and planning. rated one by one.

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

Opus 5.5

browser use
72/100
personal advice
80/100
planning
69/100

Fable 5.2

browser use
75/100
personal advice
80/100
planning
67/100

API pricing

rates, the run estimate and the bill are three different things.

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.

Opus 5.5

official API rates
Official API rates in US dollars per one million tokens
per 1M tokensinputcache readoutput
Standard$4.00$0.20$20.00

1M context window; model page lists one Standard token rate. 5m cache writes: $5.00; 1h: $8.00. Tool fees are separate; preserve dated provider estimates.

US dollars · Standard processing · Official pricing: platform.claude.com · checked

recorded estimate for this run
US$88.82partial usagePartial API-equivalent estimate: 18 of 18 saved execution records priced.method: Recorded execution usage priced at the frozen official rates. priced 8 Oct 2026.
actual bill
unknown. billing for this run: claude max subscription (both setups). no receipted charge exists for it, and unknown is not zero.

Fable 5.2

official API rates
Official API rates in US dollars per one million tokens
per 1M tokensinputcache readoutput
Standard$8.00$0.20$40.00

1M context window at standard pricing; pricing page lists one Standard token rate. 5m cache writes: $10.00; 1h: $16.00. Fast mode ($16 / $80) is not used; run on the Claude subscription, so this is a list-price comparison, not a bill.

US dollars · Standard processing · Official pricing: platform.claude.com · checked

recorded estimate for this run
US$38.71partial usagePartial API-equivalent estimate: 12 of 18 saved execution records priced.method: Recorded execution usage priced at the frozen official rates. priced 10 Nov 2026.
actual bill
unknown. billing for this run: claude max subscription (both setups). no receipted charge exists for it, and unknown is not zero.

Both setups ran on Lennox's Claude Max subscription: Claude Code for the build window and the Claude desktop app for the live lanes. Nothing was billed per token. The API-equivalent estimates price recorded execution usage at dated list rates. They are not bills, and the two do not compare like for like: the Opus 5.5 estimate covers 18 of 18 saved execution records and the Fable 5.2 estimate covers 12 of 18. Native sessions, setup, calibration, coordination, grading, presentation and subscription fees are excluded. Actual incremental spend is unknown.

review conditions

a requested blind comparison

Rated blind as Candidate A and Candidate B on the BuilderBench score deck; names were revealed after the scores locked. Model names were masked in transcripts and videos, but one planning transcript mentions a model's release date and list price.

who runs this: Lennox Saint. the v3 tasks come from real work for Threadify, the business he co-founded.

The owner rated the saved outputs and the recorded live lanes on the BuilderBench score deck. The bench played the owner in both founder conversations from saved context. Six of the eight build tasks were scored on first-round work for both setups, and the computer task counts the browser session only (see limitations). The automation repair task was scored by the BuilderBench conductor from recorded evidence. The owner authorised the partial finalization on 11 November 2026. The ratings did not change after that, and the bench recorded the lock and the reveal on 13 November 2026. Quality is the weighted task score. v1 to v3 use 70% quality, 20% autonomy and 10% proactivity. The overall score is worked out from unrounded quality, so it can sit 0.01 away from a sum of the rounded figures shown.

scores locked
yes
names revealed
yes, after lock

frozen setup

what was tested

pack version
3.1.0
build window
4 hours
reasoning
Extra High (xhigh)
speed
Standard
billing
Claude Max subscription (both setups)
scored runs
1 per setup

exact model and harness version strings are not in this public export, so they stay unknown here.

limits

what this result does not prove

  • This is one owner-finalized partial personal assessment of one run of each setup. It is not a universal ranking and it is not statistically conclusive.
  • Finalized ratings do not mean every task or recording passed verification. The original execution and its evidence are incomplete.
  • A usage limit cut off Fable 5.2's revision round on six build tasks (app lead magnet, automation repair, landing page, YouTube packaging, research, video editing). Both setups are scored on their first-round work for those six. Motion and writing use final work.
  • The computer-use lane (no browser) was not run because of the owner's time limit. The computer task counts the browser session only, for both setups. This is why the result is labelled partial.
  • Fable 5.2's planning conversation had two in-app question cards that the transcript did not capture; operator notes mark where they were.
  • Three of the four founder conversations ran one to three minutes past their limits (15 minutes for advice, 20 for planning).
  • Opus 5.5's first browser session was voided for a 13-minute recording gap and rerun. Session order alternated between the setups.
  • In Opus 5.5's browser session one permission card had no clickable buttons and waited about one to two minutes before the operator approved it; Fable 5.2's session had that fix in advance. Opus 5.5 reached the 30-minute limit without saving a file.
  • Fable 5.2's two conversation time windows were corrected after the fact; the correction is logged.
  • Both setups used the same provider, tools and settings (Extra High effort, Standard processing). An image generator was available; neither used it.
  • The API-equivalent estimates do not compare like for like: the Opus 5.5 estimate covers 18 of 18 saved execution records and the Fable 5.2 estimate covers 12 of 18. Both setups ran on a Claude Max subscription, so nothing was billed per token.
  • The score belongs to the full setup: model, harness, tools and settings.
  • Pack 3.1.0 scores compare only with other pack 3.1.0 results.
  • The build window was four hours.
  • The picture on this page is a plain score card drawn from the published numbers. The run's own reveal graphic has not been made yet.
  • In both founder conversations the bench, not the owner, wrote the owner side from saved context; the owner scored the recorded transcripts.
BuilderBench by the LAB score card: Opus 5.5 scored 74.25 out of 100 and Fable 5.2 scored 69.43 out of 100 on pack 3.1.0, 10 November 2026. Opus 5.5 leads by 4.82 points. Opus 5.5: quality 73.79, autonomy 73, proactivity 80; task scores writing 51.4, landing page 70, app lead magnet 71, video editing 78, motion 82, thumbnails 77.2, browser + computer 72, research 80, automation 89, founder conversations 74.5. Fable 5.2: quality 65.34, autonomy 77, proactivity 83; task scores writing 39.8, landing page 40, app lead magnet 57, video editing 76.67, motion 68, thumbnails 76.8, browser + computer 75, research 71, automation 89.33, founder conversations 73.5. Owner-finalized partial personal assessment of one run each. Both ran on a Claude Max subscription; nothing was billed per token.
the result graphic. owner-finalized partial personal assessment. costs are estimates, not bills. every number in it is also on this page as text.

p.s.

get the next BuilderBench result first.

  • the exact prompt i typed
  • what came back
  • where it fell over

free. unsubscribe any time, one click. privacy.

  • 34,473 on Threads
  • 3,612 readers
  • 1.91K on YouTube

live count, 22 nov 2026.