Lenny’s AI Builders

one of my work chats used 2.51% of my weekly allowance. that was the whole automation chat, not one run. in this episode, i look at my usage page, the chat-and-work merge announcement and its walk-back as recorded on 29 october 2026. then i run 12 transcripts through the API and show the output, tokens and time. the credits shown in this recording don't cover Codex. check the terms and expiry on your own account before using them. the test took 16.83 seconds. the recorded list-price calculation was about seven cents - not a promise about your bill or current pricing. also in this episode...

Claude is quietly testing new limits on some accounts. SemiAnalysis says it's about 20% more on the five hour limit, but about 20% less over the month - great for sprints, worse if you use Claude all day every week. so i checked 5 of my Max accounts across 12 days. none of them are in the test. check yours in one session: in Claude go to settings (bottom left), then the usage tab. burn about 20 points off your five hour bar, then look at your weekly. if it moved about 5 points, that's normal. if it moved 6 or 7 or more, you're probably being tested...

OpenAI just dropped GPT-6.1 Terra - a model that drives your computer, at half the price of GPT-6.1 Sol. so i let it do the admin in a made-up plumbing office on my computer: find the enquiry, book it in, write the quote. then i gave Opus 5.5 the exact same job, against the clock. both got it right, the faster one isn't the one i'd hire, and the price surprised me. also this week: Theo says the Claude Agent SDK drains your 5-hour limit faster than Claude Code (a GitHub issue a month ago measured about 1.8x per token, and...

Kimi K3.1 Pro just dropped, so i gave it and Opus 5.5 the exact same prompt: build me a 3D game where you're a Perth plumber racing to call-outs. one game cost $1.08, the other $3.21 (API-equivalent, one shot each), and only one of them got Perth. i play both blind before i tell you which model built which. also this week: OpenAI paused its best models for 11 days after a settings mistake gave one a little bit of internet during training, Cursor's agents can now drive your whole computer, Block's Dessa is the most human AI voice i've heard (waitlist only), and...

Claude Slides is now in every Claude chat, even on the free plan, and ChatGPT Slides just got live collab on Pro. so i gave both the exact same job - a 10-slide client pitch deck from one page of notes - twice each. then a free GitHub skill called frontend-slides crashed the fight in Claude Code. six decks, letters A to F, no names, and a blind judge from a different AI scored them before i looked. the loser costs me my ChatGPT plan. also this week: Meta's Muse Glimmer open model for your own computer, Anthropic's new Public...

PHP got hacked wide open this week - two critical bugs that let anyone take over a PHP server without logging in. the same day LiteLLM 1.107.0 shipped a password stealer, and Railway, where my app runs, got broken into. so i gave Claude one prompt to check my Mac and my websites for all three. PHP: clear. LiteLLM: clear. Railway: hit. Threadify runs on it, and Claude found Railway's warning email sitting unopened in my inbox. we'd already handled it, but that's the one that mattered. also this week: Claude Duet, an agent that comes to you (waitlist, US Pro...

four Shorts. two of them were made by Claude from one prompt, with my cloned voice and my AI face. pick your two letters in the comments before the reveal. if most of you can't tell, Claude makes every LAB Short for a month. also this week: Codex now keeps your chats cached forever (and why my Claude Max plan kept draining), Anthropic's draft IPO filing ($16.3 billion in revenue, and an $80 billion loss that's mostly paper), Dario's huge reply to his critic, one brain for all your agents with a lock so they stop tripping over each other, and the...

devs are paying double for the old GPT on the API - GPT-5.6 Sol costs $4 in / $20 out per million tokens, GPT-6.1 Sol $2 / $10 - and swearing it's better than the new one. so i ran both on three of my real jobs (a broken test in Threadify, a video pipeline bug and my landing page copy), two runs each, with a blind judge from a different AI that only saw A and B. also this week: GPT-6 Pro running agents in parallel, Claude writing 90% of Claude Code, X paying creators to post video, Claude Code answering sooner, and moving your AI...

Meta just gave away Muse Spark Open, which it calls its most capable US open-weight model. the catch: to run it yourself you need two Mac Studios (or rent it on OpenRouter from $1.25 per million tokens in). so i put it on BuilderBench, my own benchmark, against Meta's paid Muse Wonder Full. the result: Meta's free model scored 27.68. its own paid Muse Wonder Full scored 42.52 and Google's Gemini 4 Argon 45.68 - same tasks only (the desk tasks weren't run), rated by me. also this week: Opus 5.5 fast mode with no waitlist ($8 in / $40 out per million tokens), Claude Code Review for Team...

Claude now runs a whole team of agents - so i pointed four of them at my own landing page. it scored 46/100 on Google's speed test (Lighthouse, mobile). i gave them one prompt - Voxyz's "20 things that slow your site down" - in the new Claude Code desktop app, and let them fix it in one pass. the result: on a copy of the site with every fix, measured side by side with an untouched copy (which scored 48), 48 → 86 (median of 5 runs each). the phone freeze went from about 4 seconds to 55 ms and the page got 69% lighter (2.3 MB → 0.7 MB). the live site...

Meta and Google both say their new models caught Opus 5.5. i paid $50 for Meta's Muse Code plan, gave Muse Wonder Full and Gemini 4 Argon the same 10 real jobs from my business on BuilderBench, scored the main jobs blind, then locked and revealed the scores live. result (BuilderBench v3, 14 Oct 2026, owner-finalized partial personal review): Gemini 4 Argon 45.38/100, Muse Wonder Full 39.91/100. neither beat Opus 5.5's 68.26 - but that score is from pack 1, an older version of the test, so it's a reference line, not a same-test match. was it fair? same 10 jobs and the same 4-hour window (a storage problem interrupted it and...

X just merged Grok, Cursor and X Premium into one plan. i switched, and my plan check says it saves about A$74.60 a month. but there's a catch if you use Claude or ChatGPT inside Cursor: third-party models now bill at the provider's API price. plus Meta's Muse Wonder Full lands tied 2nd on the Artificial Analysis Intelligence Index, Hermes and OpenClaw are back on Claude Max plans, OpenAI's day 7 lifts Sol's default context from 272k to 500k, Google's AI Edge Foresight keeps your voice notes on your Mac, and Satya Nadella says to treat models as insider risks. the...

Codex now writes your next prompt. i let it run a whole session and kept 9 of 10. plus Grok Bot gets its own email, Claude Code Projects opens up, ChatGPT dots on your phone, one Claude agent that runs a thousand, Anthropic's report on Claude doing things it shouldn't, and a one-line prompt that makes your AI animations move like Pixar. the thesis: this week, agents got their own stuff. the practical: give Claude any animation and this one line: "Review this animation against John Lasseter's principles of animation and tell me the top three fixes." attach the video, or point...

OpenAI's coding model just got up to 8x faster - if you pay $500 a month. plus Claude Dashboards and Motion, a guide to Claude Managed Agents, Voyager, Firecrawl Universal Scrape, two safety stories, and i set up a Claude agent that checks my landing page, live. the thesis: they sell the shovels. we build. the practical: one command set up a Claude Managed Agent that opens my landing page, follows every button and tells me in Slack what's broken. first real run: about 90 seconds, 21 cents against a $5 cap, three problems found. this version finds and proposes only - the "say...

Elon says his own bot will use Claude. plus Claude's new Haiku 5.5, monthly API credits for Max plans, GPT-6 in ChatGPT, and i test Haiku 5.5 against GPT-6 Luna on BuilderBench, live. the thesis: competition is great - builders are cashing in. the practical: i ran BuilderBench v3 on Haiku 5.5 and GPT-6 Luna (8 Oct 2026, 4 hours, extra high) and rated every job blind before i knew which model made it. GPT-6 Luna 44.58, Haiku 5.5 39.18 (out of 100, final personal scores, partial evidence). run cost estimates: Luna $1.47 (short-context scenario), Haiku $37.39 (provider-reported). these are API-equivalent estimates, not bills, and the methods differ. one run, my jobs...

OpenAI promised 28 days of gifts. on day two, poteto - Lauren Tan from the Grok Bot team - called them nothingburgers. there is beef. 76% of almost 75,000 votes on Tibo's poll said day two needed a reset. we got one. my take: build the model, not the calendar. then Claude moves into Google Docs, Sheets and Slides (beta, paid plans), Hark Pro (the Figure founder's personal agent), OpenAI's AI did new maths (with a model we can't use yet), treg vs Clay for lead gen, and Nano Banana 2.1. the practical: the one prompt that makes Claude use every drop of my...

Claude gives you about 5x more for your money than OpenAI. that's not my maths - SemiAnalysis measured it. they ran the same jobs on every big AI plan and watched the usage bars. on the $200 plans, Leo's screenshot of their data shows 28.6 billion tokens of Opus 5.5 vs 10.2 billion of GPT-6.1 Sol. that's 2.8x the raw tokens, and about 5x the value at API prices. one lab even gave 1 of 3 identical subscriptions 20% lower limits. so know what you're paying for. we deserve to know. then OpenAI's 28 days of gifts (my take: an advent calendar for people who just lost half...

i think the AI companies built a casino. OpenAI keeps refilling everyone's Codex usage at random. one tracker has counted 58 resets. people bet real money on Polymarket on how many come each week. then a reset promised for 10am landed about four hours late - after one guy burned his last 30% so he wouldn't lose it. so i break down the dark psychology behind it: Skinner's pigeons from the book Hooked, why the wait is the hook, and why losing hurts more than winning feels good. it's probably not that deep. but i don't think you get that big without...

people are asking for Fable 5.1 and getting something smarter back. Anthropic hasn't said a word. so i ran the test myself, in Claude Code. plus Claude Code mods, half-price Claude design, Karpathy on visual outputs, Anthropic's secret dinners (a report), putting your ChatGPT dot in charge, and the practical: i had Claude Code read 8 episodes of my YouTube analytics and it found where you stop watching. watch the full episode: https://youtu.be/nVZf_vim5K4

ChatGPT wants your whole day. it now takes your meeting notes with no bot on the call. it hosts apps you build with one sentence. so i let GPT-6.1 Sol audit my life and design my ChatGPT Space. it worked for 52m 54s. Claude Opus 5.5 checked the plan over four rounds until it said AGREE. what came out: one "Lennox OS" Space where i capture, plan, start work, review results and find my history. the hookups to my other apps aren't built yet - i show you exactly where it stands. also in this one: Google's Gemini 4 Argon can write 1...

OpenAI's new $500 plan feels like the old $200 plan. that's one of six stories from OpenAI DevDay 2026 - then i set up my own dot live, on day one, with the first prompt every dot needs. the stories: the new $500 Pro plan (25x the usage of Plus, plus Ultrafast - up to 8x faster, and $60/$300 per million tokens for Astra in the API) and the $200 plan dropping from 20x to 10x Plus (Tibo: it will "net out at half the dollar in API spend"), GPT-6.1 Sol at $2/$10 per million tokens vs GPT-6 Astra's $10/$50, dots (always-on GPT-6 Astra agents with their own...

Anthropic wants to be worth more than $2 trillion, and its own filing warns its AI could act in ways resembling blackmail. that's one of six AI stories this week - then i gave Sonnet 5.5 four hours of my real work on BuilderBench, the benchmark i built for builders, and priced every run. recorded an hour before OpenAI DevDay. the stories: Anthropic's IPO filing, per Reuters, who saw the filing (nearly $4.6 billion in revenue last year, $11.5 billion from April to June this year, more than $8 billion lost running the business last year, $518 billion promised for compute and cloud, the founders keep 50.1...

Opus 5.5 leads the Artificial Analysis chart. GPT-6 Sol has a lower published API list price. The half-price comparison refers to standard short-context API input and output list rates, not subscriptions or cost per finished task. But which one should you use for your own work? i look at the launches, the chart, cost per task, and one Sol miss in my own Notion planning. Then i show how to set up a five-task test for the jobs you actually do. i have not run Opus 5.5 or scored BuilderBench yet, so this episode does not pick a real-work winner. try this...

i looked at OpenAI's reported maths results, the GPT-6 Sol and Opus 5.5 price fight, Grok 4.7, Muse with Shopify and Qwen-Image-2.1. Then i gave Astra my existing LAB brand and asked for a matching thumbnail, banner and podcast cover. Astra directed the work in Codex; native ImageGen made the images. I show the finished set, the references and the checks. Give it your brand first. The maths result is OpenAI's claim, not an independently established result here. Some model details were reported or rumoured when this was recorded. Qwen's downloadable model needs suitable hardware. Threadify sponsors this episode. It helps turn...

OpenAI and Anthropic are looking into tens of thousands of times their models went off script. that's one of six AI stories this week - then i scored GPT-6 Astra on BuilderBench, the benchmark i built for builders, and tracked what it cost. the stories: Axios reports OpenAI, Anthropic and security researchers are investigating tens of thousands of incidents where frontier models did things outside evaluators would call problematic (unnamed sources - labs run hundreds of thousands of tests, and most aren't known to have caused real harm), the OpenAI DevDay countdown and what leaked from OpenAI's code (an always-on...

six AI stories this week, then a test: i gave Claude Opus 5.5 the same prompt at every effort level - low, medium, high, xhigh and max - and tracked the time and cost of each.

the stories: OpenAI says its research agents used misaligned strategies about two dozen times in training and testing (most low severity, review still going), Vercel's AI Gateway token spend shifting toward OpenAI and open-weight models, a reported $500 ChatGPT Pro Max plan (not announced by OpenAI), Claude's new plugin portal, Claude Code stopping gracefully at your five-hour limit, and the UK's draft advice...

four AI stories this week, then a blind test. i hid the model names, made GPT-6 Sol and Claude Opus 5.5 do the same eight builder jobs, scored the work, and only then revealed who was who.

the stories: ChatGPT voice can now use plugins, an agent developed by OpenAI got into an Australian government statistics portal (still under investigation, with no confirmed personal-record exposure), Meta's new Muse personal agent (Muse Charm was shown, not shipped yet), and Claude finding a new enzyme system (early days, and what it does is still unknown).

the result: Opus 5.5...

i got tired of giving my AI the same feedback. in this episode, i show the thumbnail changes that led me to make a remember-this skill. first, the news: the Codex reset request, Jev browser clicks, CUA-S1 forms, moving skills between AI apps, handoff files, Claude's role in AI research, and Astra for Law.

watch the full video for the screen demos and visual examples: https://www.youtube.com/watch?v=1S7eR8VSgVI

grab my free goal skill: https://github.com/lennoxsaint/Skills/tree/main/goal

recorded 20 September 2026. the reset discussion reflects...

six AI updates, my take on each, and the Email Triage Pro workflow i keep pinned. recorded 19 September 2026. offers, pricing and account access may have changed since recording. the Pro 20x story is a limited-access report, not an official general reopening announcement. field notes: https://notes.lennoxsaint.com/posts/lab-0003 video: https://youtu.be/6ABasoGZY9Q

OpenAI published a framework for tracking, investigating and disclosing model misalignment. Claude is merging chat and Cowork. Grok Bot says its 1Password connection asks before each fill. then i show how one voice dump became two focused Codex threads. get the full field notes, every source and the copyable voice-conductor prompt: https://notes.lennoxsaint.com/lab-002 in this episode: - what OpenAI's model-misalignment framework says and what it does not prove - Claude bringing chat, Cowork, docs, slides and design into one place - Grok Bot using 1Password with approval before each fill - Nick Baumann's GPT-6 Astra motion-design workflow ...

Sam Altman's OpenAI teaser, early Opus 5.2 reports, Grok Bot's three-day build, Jev, Persona and Zuckerberg on AI safety - then the practical Codex workflow Lennox made from feedback he kept repeating. The teaser does not confirm a launch. The Opus impressions are one user’s report, and Lennox has not measured time saved by his own workflow. Sources: Sam Altman on OpenAI releases: https://x.com/sama/status/2099872600977760451 One user’s Opus 5.2 report: https://x.com/pankajkumar_dev/status/2099892088368476668 Grok Bot three-day build, Day 1: https://x.com/bot/status/2099890276315848743 Jev introduction: https://x.com/CompleteSkeptic/status/2099925682726002904 Persona wearable: https://x.co...

Who gets to decide how fast frontier AI moves?

In episode zero of Lenny's AI Builders, Lennox looks at Dario Amodei's proposal to pace the frontier, Barack Obama's call for a wider public role, and Jack Dorsey's argument for an open frontier. Then he turns it into a builder question: can small builders still get a shot?

The practical takeaway: if your agent keeps building scaffolding, say what the job is, name the finished result and deadline, and ask for the simplest path.

Watch LAB 0000

Read the LAB 0000 field notes

Dario Amodei has proposed pacing frontier AI development. i read the key passages, explain the three-step plan, distinguish Sam Altman's concrete commitment from the wider proposals, and talk through my own mixed feelings as someone who loves building with AI. this is an unnumbered LAB pilot. for audio listeners: the three “here” call-outs refer to separate passages discussing what extra time could help researchers achieve. they are not an announced model-release freeze. read the essay: https://darioamodei.com/post/we-must-pace-the-frontier watch the video: https://youtu.be/TFS7Ef6_7oo Lo-Fi Beat | Coffee by Alex-Productions | https://onsound.eu/ Royalty Free Music by h...

I tried using Astra to edit in DaVinci Resolve. This audio adaptation covers why I wanted the workflow, why I record camera and screen separately, how Gemini entered the review process, and where the agent stalled or stopped for a review decision. The video tutorial contains the setup screens and examples. This version keeps the spoken experiment and tradeoffs, removing passages that need visible price and usage corrections. Explicit language. The overnight story describes one experiment; it is not a promise of unattended completion. Watch the full tutorial: https://youtu.be/D28EiNzw52c

this is the unchanged soundtrack from my Astra–Fable video-editing comparison. the screen recordings, edits and scorecards are in the video:
https://youtu.be/fYuok6zOlPo

i gave GPT-6 Astra and Fable 5.1 the same video and two rounds of prompts. each made three Shorts and three long edits. then i added my own long edit and scored all 13.

i want to hand an AI a recording, go away, and come back to a finished video. so i checked how much nudging each model needed, then watched what it actually made.

watch the ed...

OpenAI says its next-generation internal model helped solve the Navier-Stokes Millennium Prize Problem. Around 10,000 coordinating agents. 88 hours. I wanted to understand what that means without needing a maths degree.

I walk through the announcement, the vortex, and a plain-English explanation from GPT-6 Astra. Then I give my take on what this could mean for AI and scientific discovery.

Watch the video: https://youtu.be/LlAoO5AErJU

CHAPTERS

0:00 OpenAI’s new maths claim

0:29 The internal model beyond GPT-6 Astra

1:09 10,000 ag...

i let ChatGPT-6 Astra take over my computer for five demos, then gave it a bonus test: edit a YouTube intro in Descript. This is the audio from the hands-on video. Music, Google Earth, Keynote, Photopea and Blender all get a turn. The editing test shows the difference between using an app and matching my taste. 00:00 my first reaction 00:43 make a soundtrack 01:19 scout locations in Google Earth 01:44 animate a Keynote presentation 02:31 design a poster in Photopea 03:08 build a 3D sculpture in Blender 04:47 bonus: can it edit my YouTube video? 08:13 my verdict Watch the screen recordings: https://youtu.be/wFBO36lOVpw

i told ChatGPT to inspect one month of recurring charges and keep working until it found a real ten-percent cut.

it had one restriction: both OpenAI accounts and any service still doing an essential job stayed.

this episode shows the receipts, the seven-step Save 10% skill, where the agent had to stop for approval, and why a projected saving does not count until the provider proves the next bill is lower.

chapters:

After two and a half years across Saints College and Full Circle, I'm closing this chapter. In this short personal episode, I explain why running a community on Skool pulled me away from the Threads community I was helping others build, why I couldn't keep doing both, what Full Circle gave me, and why I'm going back to basics. This isn't the end of the book. It's the end of the Full Circle chapter. See you on Threads: https://www.threads.com/@lennox_saint More from me: https://lennoxsaint.com

ChatGPT observed every minute of one workweek, producing 12,421 images. After removing 170 exact duplicates, Codex analysed 12,251 unique frames for repeated work before the final judgment point.

The audit returned seven findings and two honest totals: Codex found 5.02 hours; Lennox accepted 4.21.

That gap is the useful part. The model's opportunity total and the owner's decision are allowed to disagree.

Chapters
00:00 ChatGPT watched every minute
00:41 What Chronicle does
00:53 The five-hour bet
01:28 12,251 unique frames
02:08 The rules and fourth analysis
03:15 Seven findings begin
05:47 Cross-thread work
06:12 Video production loops
07:31 5.02 found...