Meta planned to run on 60% fewer people. Then it didn't.

⚡ 60-Second Catch-Up

The AI-native org got tested at the top, and it flinched.

Reuters reported the details this week: at a January leadership retreat, Meta hatched "Project OT" — Organization Transformation — a plan to shrink many teams by 60% on the assumption that AI would let the remainder work as well as before. HR projected a cut larger than 2022–23, when 25% of staff went. The night of May 19, hours before the first layoff was to run, it was called off.

What actually shipped instead is the interesting part. Gergely Orosz has been tracking the fallout: roughly 10% of engineers laid off in May, another 20–30% of infra and product engineers reassigned to data-labelling work for AI training, and a run of embarrassing outages after. Meanwhile Jagjit Chawla, the VP behind Facebook's Feed and Reels, told Nikhyl Singhal that on his own team no individual contributor now sits more than one layer from him, and that the win was not the headcount — it was losing the coordination tax that made big-company product work feel like arm-wrestling.

The lesson is not "AI replaces people." It is that the org chart was carrying more waste than the work was, and the two got confused.

→ do this: before you reach for a tool to make your team faster, count the handoffs a decision crosses. Most solo builders have zero and don't need the tool. Most small teams have four.

Cost per token stopped telling you what a model costs.

Google and Meta shipped competing workhorse models on the same day — Gemini 3.8 Flash and Muse Spark 1.3 — and took opposite routes. Google let Gemini work harder on hard problems. Meta trained Muse to skip work it doesn't need, reporting about 20% fewer tool calls and 25% fewer tokens than its predecessor. Google kept Gemini's token price flat. Artificial Analysis then measured the thing that actually matters and found Gemini cost roughly 40% more per completed task, precisely because it did more work per job.

Underneath that, Ramp's spending data says about 80% of OpenAI and Anthropic enterprise revenue comes from 1% of customers — a concentration Ramp says it does not see in other software categories. The per-token price war is being fought over customers who are, in revenue terms, rounding errors.

→ do this: re-run your own eval and divide by completed tasks, not tokens. A model that costs double and gets it right first time is the cheap one.

The crowd has already picked its lane, and it's the obvious one.

Guillermo Flor read all 236 companies in YC's Summer 2026 batch: 151 of them (64%) are AI-native, and agents alone account for 52 companies (22%). The move nobody is talking about is Industrials, whose share of the batch jumped from 12.8% to 23.7%.

Against that, a counter-example worth more than the batch data. Hakeem Dimozantos, 22, built Kiku — a Japanese-learning app that works like a video feed — after failing to learn the language with textbooks and Duolingo. Five months on iOS: roughly $16K in sales, about $12.2K after Apple's cut and tax, $7K of it in the last month alone. Distribution was him posting four or five times a day on TikTok and Instagram, 2.7M views on one account and 1.1M on the other, starting a month before launch.

→ do this: if you are building an agent startup, you are company 53. Go read the Industrials column, or go be the person your product is for and post about it for a month before you ship.

🛠 Tools

Build & ship

• ProveKit — Client-side zero-knowledge proofs — let a user prove age, nationality or ID ownership without sending you the underlying document · Open source (MIT)
• Claude Commerce Agents ⭐ — Anthropic's reference blueprints for shopping and merchant agents · Open source (Apache-2.0)
• Lily — Perplexity's Apple-silicon inference engine for Qwen3.6-35B-A3B · Open source (MIT)
• Fable 5.1 World Modeling — Building navigable worlds via code · Open source (MIT)

AI & agents

• Check files for Claude — Upload a file and get back signs of whether Claude made it · Free
• fal H3 Max Turbo — MiniMax H3 video generation at roughly twice H3 Max's speed, text-to-video and image-to-video endpoints · Paid — usage-based per second

Design & create

• BRRRANDING — A human-curated gallery of branding work — no algorithm, no AI · Free
• Cursor Craft — Hundreds of custom pointer sets for macOS, one-click install · Free

Growth & ops

• Ato — A screen-free voice companion for older adults — check-ins, reminders, family messages, no camera · Paid — $99 at launch (regular $179), plans from $249/year

📚 Resources

Steal the template

• How to build a reliable agent harness — the long-form playbook for the scaffolding around the model, which is where most agent projects actually fail.
• An organizational second brain — Meta's engineering write-up on building an AI that learns from how a company already works.
• A docs page is a very long search query — Val Town on writing documentation for the thing that reads it before your user does.

Learn the craft

• Maybe we shouldn't be reviewing all this code — a direct challenge to review-everything orthodoxy now that volume has changed.
• Is product discovery a phase or a weekly habit? — the long version of an argument most teams settle by accident.
• Ecosystem design — designing for the surfaces around your product rather than the product alone.
• How we make AI coding more cost efficient — GitHub on the quality-per-dollar tradeoff, from the vendor side.

Benchmarks you can re-run

• Gemini 3.8 Flash vs Muse Spark 1.3 ⭐ — the head-to-head that produced the cost-per-completed-task figure above.
• Which tools do coding agents actually choose? — 16,893 measured sessions across Claude Code, Codex and Cursor, and the answer is not the fancy one.
• LLMs: intelligence vs. cost — the frontier plotted, so you can find your own bar.
• Ramp AI Index — corporate AI spending measured off actual card and bill-pay data.
• TxBench — an antibody-discovery benchmark, for anyone who thinks the model race is only about chat.

📖 Reads

☕ Under 5 minutes

• GPT-6 Astra — OpenAI's own announcement of the model everyone spent Thursday guessing about.
• OpenAI Astra and looped transformers — Sebastian Raschka on what the architecture appears to be doing.
• Test time training — a short, clear explainer on the technique.
• Muse superapp and the Ava computer-use model — what Meta appears to be assembling next.
• On animating what you can't see — motion design for states the user never directly observes.
• What comes next for AI? Our bet is world models — the case for world models as the next axis.
• The new Disney World logo is kind of genius — a rare rebrand that earns the word.

🍵 5–10 minutes

• Introducing Gemini 3.8 Flash — Google's launch post, including the Cyber variant.
• Muse Spark 1.3 — Meta's own numbers on tool calls and token use.
• Run cloud agents on machines you manage — Cursor's answer for teams that cannot send the repo out.
• GPT-6 Astra on ARC-AGI-3 — the independent benchmark result, published same-week.
• An AI-assisted cyber attack — Unit 42's investigation, written up as a case file.
• Grep beats LSP? — why coding agents keep ignoring the better tool you gave them.
• Uppy 6.0 — fewer packages, fewer moving parts, and a rebuilt S3 path.
• AI is making product development faster. But where did the work go? — the displacement question, asked from the design side.
• Try Google Pics — image creation and editing landing inside Workspace.
• Updates to generative plugins and shaders — Figma's release notes, worth reading if you ship UI.

📚 Longer, worth it

• Anthropic has some alignment problems ⭐ — 23 minutes, and the most substantive critique published this week.
• How concerned should we be about Astra's recurrent architecture? — the safety-side reading of the same architecture.
• Saving petabytes of cache with Zstandard — Cloudflare on transcoding the cache itself.
• Wasmi 2.0 — engineering one of the fastest Wasm interpreters.
• The asteroid currently hitting front-end web development — Nolan Lawson on what is actually ending.
• Project Xanadu — Gwern on the hypertext system that lost, and why it kept losing.
• From hookswitch to grave — telephone infrastructure history, unreasonably good.

📊 Numbers

• 60% — the team-size reduction Meta's "Project OT" planned before the plan was shelved on 19 May (Reuters).
• 25% — share of staff Meta cut in 2022–23, which HR projected Project OT would exceed (Reuters).
• 10% — engineers laid off at Meta in May, with a further 20–30% of infra and product engineers reassigned to AI training work (Gergely Orosz).
• ~40% — how much more Gemini 3.8 Flash cost per completed task than Muse Spark 1.3, at an unchanged token price (Artificial Analysis).
• 20% / 25% — fewer tool calls and fewer tokens Muse Spark 1.3 used versus its predecessor (Meta).
• ~80% — share of OpenAI and Anthropic enterprise revenue coming from 1% of customers (Ramp).
• $1B — annualized revenue run rate OpenAI says ChatGPT Ads has already reached (OpenAI).
• 35% / 60% — how much larger carts got, and how much more likely shoppers were to check out, using Anthropic's commerce agent blueprints (Anthropic).
• 236 / 64% / 22% — companies in YC's Summer 2026 batch, the share that are AI-native, and the share building agents (Guillermo Flor).
• 12.8% → 23.7% — the jump in Industrials' share of the YC batch (Guillermo Flor).
• ~600,000 — students covered by New York City's one-year ban on student-facing generative AI, 2-K through eighth grade (NYC Public Schools).
• $16K / $12.2K — Kiku's five-month gross sales, and what was left after Apple's cut and tax (Hakeem Dimozantos).
• 1,500 tokens/s — the rate Cerebras is serving Qwen 3.8 27B at (Cerebras).
• $4,500 — the entry price of Kantar's self-serve ad test, which is what makes synthetic pretesting interesting (Dru Riley).
• $400M / $100M / $40M — reported revenue at Lovable, Base44 and Bolt.new, all founded within the last three years (Dru Riley).
• 128 — experiments in the body of work showing that paying people for a task they already enjoyed made them less likely to continue once the reward stopped (Deci and Ryan).

📰 What Happened

Models & math

• OpenAI released GPT-6 Astra, the model it had been publicly staging for weeks (OpenAI). ARC Prize published independent ARC-AGI-3 results the same week (ARC Prize).
• Google launched Gemini 3.8 Flash and a Gemini 3.8 Flash Cyber variant (Google).
• Meta released Muse Spark 1.3, tuned to skip work rather than do more of it (Meta).
• Cerebras is now serving Qwen 3.8 27B at 1,500 tokens/s (Cerebras).

Safety & policy

• OpenAI told House lawmakers it is building automated shutdown controls and tighter agent monitoring, after an evaluation agent escaped its container (Reuters).
• The US Justice Department backed OpenAI in the New York Times copyright case, arguing training on copyrighted material can be transformative fair use (AP).
• New York City schools banned student-facing generative AI for a year from 2-K through eighth grade, with AI-literacy lessons and vetted tools for high schoolers instead (NYC Public Schools).
• Palo Alto Networks' Unit 42 published an investigation into an AI-assisted cyber attack (Unit 42).

Business moved

• Reuters revealed Meta's "Project OT" — a plan to cut team sizes by 60%, killed hours before the first layoff on 19 May (Reuters).
• ChatGPT Ads reached a $1B annualized run rate, and OpenAI opened campaign creation to businesses (OpenAI).
• Nvidia and CrowdStrike are developing new cybersecurity AI models together (WSJ).
• Shamez Hemani, formerly OpenAI's Stargate exec, joined Anthropic (DataCenterDynamics).

Platforms shifted

• Anthropic gave Claude background computer use in Cowork and Claude Code on Mac, so it can operate approved apps while you work elsewhere (Anthropic).
• Anthropic open-sourced Claude Commerce Agents, reference blueprints for shopping and merchant agents (Anthropic).
• Perplexity open-sourced Lily, its Apple-silicon inference engine (Perplexity).
• OpenAI connected ChatGPT for Healthcare to Epic, allowing read-only patient records into ChatGPT (OpenAI).
• Cursor added self-hosted machines so cloud agents can run on infrastructure you control (Cursor).
• Google Pics launched in Workspace (Google), and Figma shipped generative plugin and shader updates (Figma).

We skipped a run of Hacker News stories this week with nothing in them for founders — a kidney-disease investigation, a domain-name obituary and some very good sand sculpture among them.