neither caught Opus

from the episode: Meta and Google both say they caught Opus - so i tested them | LAB 0021

Lenny's AI Builders - LAB 0021

Gemini 4 Argon 45.38, Muse Wonder Full 39.91 - same 10 real jobs, scored blind.

Meta and Google both say they caught Opus. i think that's bullshit.

so i paid $50, gave both models 10 real jobs from my business, and scored them blind, live.

if either one beat Opus, i said i'd admit it right here. neither did.

why it matters

Google's Gemini 4 Argon went public on 13 October. Cursor says it's on par with Opus 5.5. a day earlier, Meta shipped Muse Wonder Full and said it beats top models on long agent work.

a few months ago, Meta wasn't even in the race. now both say they're level with Opus - the model most builders pay for. if that's true, what you pay for changes.

i don't trust press releases. i trust my own test, on my own jobs.

what you get

a real answer from real work - not a launch post.

a winner. Gemini 4 Argon beat Muse Wonder Full, 45.38 to 39.91 out of 100.
where each one breaks. Argon writes well but failed the video job. Muse Wonder Full never used the browser.
a test you can copy. the same blind setup works on your own jobs, with the two models you're choosing between.

the real example: BuilderBench, scored blind

BuilderBench is my own test. it has 10 real jobs from my business: writing, a landing page, a lead magnet app, video editing, motion graphics, YouTube thumbnails and titles, computer use, research, automation and conversation.

Muse Wonder Full ran in Muse Code, on my $50 a month plan - it has no API yet. Argon ran on Google's paid API. both got the same 10 jobs and the same 4-hour window.

i didn't know which one was A and which was B. i rated all 56 outputs, locked the ratings, then revealed the names live.

Gemini 4 Argon: 45.38. Muse Wonder Full: 39.91.

LAB 0021 frame at 11:48: the reveal cards, Muse Wonder Full 39.91 and Gemini 4 Argon 45.38, above the line "Opus still runs this town." — source: LAB 0021 episode · 11:48 - the reveal

where it was won and lost:

writing went to Argon. it scored 50 on the writing job. Muse Wonder Full scored 40. Opus 5.5 scored 49 there, on the older test. the rumors are true - Argon is a really good writing model.
Argon failed the video job. it didn't edit the long video or the Short, so it scored zero there.
Muse Wonder Full never touched the browser. almost 10 minutes in, i'd seen no browser activity. it scored zero on the computer-use job.

Argon was also good to talk to, and its thumbnails were pretty good. Muse Wonder Full was mid.

neither beat Opus 5.5's 68.26. Opus still runs this town.

try it

run your own small blind test before you switch models.

pick 3 to 5 real jobs. use work you do every week - a post, an email, a landing page. write each one as a short brief.
give both models the same brief. same files, same time limit. write down the app and the effort setting each one used.
hide the names. save each result as A or B, and let someone else - or the prompt below - keep the key.
score before you know. give every result a score out of 100 on the same sheet. then lock it - no changes.
reveal, then add the cost. write down what each model cost, or "not recorded" if you can't tell.

copy this prompt

paste this into a chat with a third model - not one of the two you're testing.

Lenny’s AI Builders / LAB 0021

i'm testing two AI models on my real work. i will score them myself.

job: [what i asked for]
result 1: [paste]
result 2: [paste]

1. remove anything that shows which model made each one.
2. shuffle them and label them A and B. don't tell me which is which.
3. make me a score sheet: 3 to 5 things a great result must do, each scored out of 100, with one line on what 100 looks like.

do not score them. only tell me who A and B are when i type "reveal".

you'll know it worked when

you scored every result before you knew who made it, and you didn't change a score after the reveal.

you can name which model wins which job, not just who won.

a bad result: you guessed who was who while scoring, you changed a score after the reveal, or one model got more time or better tools and you didn't write it down.

limits and dates

this is one run of BuilderBench v3, on 14 October 2026, with one blind rater - me. it's an owner-finalized partial personal review: all 56 ratings are final, but 3 of 4 recordings have gaps.

was it fair? partly.

same: the 10 jobs and the 4-hour window. a storage problem interrupted the run, and the clock was not reset.
not the same effort: i asked for the top setting, and Argon's highest setting is high.
not the same tools: Muse Wonder Full worked in the Muse Code app. Argon worked through BuilderBench's own API setup.
not all blind: only the main jobs were scored blind. autonomy and proactivity were rated by name.
cost is not a bill: Argon's cost is an API-equivalent estimate of about $8.92, with partial coverage. Muse Wonder Full's usage wasn't recorded - it ran on the $50 a month plan.
practice data: the research and automation jobs use labelled practice files, not live business data.

Opus 5.5's 68.26 comes from pack 1, an older version of the test. treat it as a rough line to beat - it wasn't the same test. i still need to rerun Opus on the new, stricter version.

LAB 0021 frame at 12:00: every BuilderBench score on one board, with Opus 5.5's 68.26 circled at the top and Gemini 4 Argon and Muse Wonder Full lower down — source: LAB 0021 episode · 12:00 - every score

Argon costs $2 in and $10 out per million tokens until 30 November, then $4 and $20. prices change fast - check Google's pricing page.

LAB 0021 went public on YouTube on 15 October 2026 at 8:00 pm Perth time.

the rest of this LAB

REA takes apart any app. REA - reverse engineer anything - is a free, open-source tool. plug it into Claude Code or Codex, and your agent can take apart websites, Mac apps, Android apps and firmware, then explain how a feature works. it had 186,000 GitHub stars when i filmed, up about 40,000 in 24 hours. use it on software you own or are allowed to examine. my take: closed source isn't dead. secrets are.
Grok Bot sold a $500 ad. Alex Finn says his Grok Bot built a small business on its own - a site of AI agent use cases - kept it running, then sold a $500 ad. that's his number, and he ran the social media himself, so call it half autonomous. my own agent, LennyClaw, made $0 on OpenAI in May and got no replies on Grok Bot in August. on 9 October Claude Code rebuilt it, and it sent 15 emails on day one. my take: bots are making money now.
Tibo's day 9: Codex on disk. Tibo says ChatGPT and the Codex desktop app should now store about 60% less on disk. mine dropped to 180 GB, down from 230 GB. that's not 60%. hopefully you get the full 60.
Claude broke git at Anthropic. Anthropic says it merged about 2,500 code changes a day in May and 30,401 a day by September - 12 times more in four months. so it built its own version control system and moved its monorepo off git in five months. those are Anthropic's numbers, and git is fine for you. my take: the AI isn't the bottleneck anymore. you are - reviewing the work still takes us the most time.
DigUp searches your Mac by meaning. DigUp is a free, open-source Mac app built on Google's tiny EmbeddingGemma 2 model, and Google's Gemma team shared it. it searches what's inside your files, not just their names, and everything stays on your Mac. i typed "a red circle around the words worked for nine minutes and 57 seconds", and it found it. my take: local AI is coming for Spotlight.
a bug planted in OpenAI's office. OpenAI found a hidden recording device in one of its offices, used its own models to work out where it came from, and traced it to an employee. the FBI made an arrest. charged, not convicted, and OpenAI says it doesn't look like anything was taken. my take: lock it down - insiders are a risk too.

sources

next step

watch LAB 0021 on YouTube - the practical starts at 9:17.

want BuilderBench as a free, open-source repo? comment Bench under the video. if enough people want it, i might build it.

want the next one? get the next field notes. optional - these notes stay free to read.