1 / 11
← → navigate · scroll / swipe · F fullscreen
Episode 28 · Friday, September 4, 2026 · 4:00 PM ET

Weekly Claw

The agreement landed. The models answered.
Twelve cards, three grids: Nvidia made its Hugging Face acquisition agreement official; Anthropic shipped two safeguard tiers; OpenClaw returned with 2.0; Qwen refreshed Max; New York paused classroom AI; Visko streamed living worlds; OpenAI launched GPT-6 Astra; Cerebras made Qwen3.8-27B absurdly fast; and World Labs turned video generation into spatial world modeling with Atlas.

Hosts: @AndyML · @HiM ~38 min · twelve cards · faster cuts · built to be clipped
Sponsored by Herald Labs Heritage Telecom
“The agreement landed. The models answered.”
Weekly Claw #28
01 / COLD OPEN
Cold open · the frame

The agreement landed. The models answered.

A chip company agreed to buy the open-weights home, two model tiers arrived with different guardrails, and an open-source agent platform came back from seven quiet weeks. Then New York City paused classroom AI for a year, Visko tried to stream living worlds, and World Labs put camera control and 3D reconstruction into one world model.

1
The deal
2
Model tiers
3
Platform pace
4
Public pushback
5
Policy debate
🤝Nvidia signed the Hugging Face acquisition agreement; the open-weights commons may soon have a landlord.
⚡Anthropic shipped Fable and Mythos 5.1; long agent runs got cheaper.
🌱OpenClaw came back as 2.0; the operating layer is moving fast.
Weekly Claw #28
SPONSOR
Brought to you by

An applied AI product lab where humans and agents build together. Entity is mission control for agent teams. Hacker houses worldwide.

Build with humans. Ship with agents.
labs.theherald.co
Weekly Claw #28
WHAT HAPPENED THIS WEEK
What happened this week · part 1

The agreement landed. The frontier answered.

Four launches and confirmations moved the stack from the commons to the agent shift. Each card links to the primary receipt.

The Register article capture: Nvidia buys Hugging Face for $12.9B

Nvidia signs the Hugging Face agreement.

Definitive agreement signed; about $13B, expected to close in H1 2027 pending approvals.

theregister.com
Anthropic benchmark table: Fable 5.1 versus Fable 5, Opus 5, and GPT-5.6 Sol

Anthropic launches Fable 5.1 and Mythos 5.1.

One model, two safeguard tiers; cache reads are 75% cheaper.

anthropic.com
OpenClaw release capture: v2026.8.1, AKA OpenClaw 2.0

OpenClaw ships version 2.0.

Seven quiet weeks, then v2026.8.1 touches the whole platform.

openclaw.ai
Qwen official benchmark capture naming Qwen3.8-Max-0902

Qwen upgrades Qwen3.8-Max to 0902.

2.4T parameters, 1M context, same price; post-trained for Coding and Cowork.

@Alibaba_Qwen
Weekly Claw #28
WHAT HAPPENED THIS WEEK
What happened this week · part 2

Then the world pushed back.

Schools set a boundary, video became a place, the stars aligned into a launch, and inference got stupid fast.

Chalkbeat article capture: NYC's new school AI ban explained

New York City pauses student-facing AI.

One-year moratorium through eighth grade; about 600,000 students are affected.

chalkbeat.org

Visko launches Orbis 1.0.

A Live Model for interactive worlds, with a $10M pre-seed round reported.

▶ LIVE · @viskoai launch
GPT-6 Astra benchmark table versus Claude Fable 5, Claude Opus 5, and Gemini 3.8

OpenAI launches GPT-6 Astra.

Largest run yet on 100k+ GPUs; Brockman: “Welcome to the AGI era.”

openai.com

Cerebras puts Qwen3.8-27B at ~1500 tok/s.

Shared Tier rollout; Henry wired it into ada the same day and posted the rate limits.

▶ LIVE · @imnotchalk demo
Weekly Claw #28
WHAT HAPPENED THIS WEEK
What happened this week · part 3 · the ones we nearly missed

The launch pedal stayed down.

Four more shipped while we were watching Astra: Google’s third Flash in six weeks, Meta’s agent workhorse, fal’s fastest frontier video model, and World Labs’ omni model for camera-controlled video and 3D worlds.

Google benchmark capture: Gemini 3.8 Flash and Flash Cyber

Google ships Gemini 3.8 Flash + Flash Cyber.

Third Flash in six weeks; same price as 3.7. Flash Cyber does vulnerability detection for trusted defenders.

blog.google
Meta benchmark capture: Muse Spark 1.3 agentic-workflow results

Meta releases Muse Spark 1.3.

Agentic and coding gains; about 20% fewer tool calls and 25% fewer tokens than 1.2 in Meta comparisons.

research.meta.ai
Artificial Analysis benchmark capture: fal H3 Max video rankings

fal launches H3 Max.

Post-trained MiniMax H3; #1 image-to-video on Design Arena; five-second clip in about three seconds.

▶ LIVE · @fal benchmark

World Labs launches Atlas.

One omni model for camera-controlled video, sparse-view 3D reconstruction, and robotics simulation.

▶ LIVE · worldlabs.ai
Weekly Claw #28
03 / SIGNAL FROM OUTSIDE · 6 MIN
Signal From Outside · weekly video review · permanent anchor · 6 min

Fable is a supervision problem.

THIS WEEK’S VIDEO · THE NEURON

“Claude Fable 5.1 LIVE: Testing Anthropic’s New AI Agent”

60:02 · a live test of the new model. Start around 01:22 for the model split, 03:17 for the cache-read price, 06:55 for benchmarks, 08:31 for computer use, and 11:18 for Cat Doom. The useful story is not another score; it is an agent that keeps working after the prompt.

▶ PLAY · youtu.be/9F_uP0_bTYo
The Neuron video thumbnail for Claude Fable 5.1 live test

Manual start/end cues only. No autoplay. Poster is a fallback still.

Weekly Claw #28
04 / HOT TAKE · 3 MIN
Hot take · the fight · two sides, one verdict

NYC’s one-year AI moratorium is a mistake.

One-year moratorium on student-facing generative AI through grade 8. About 600,000 students. Roughly 40 tools / 38 contracts. High-school modules and five supervised pilots remain. Is the pause the right call?

HENRY · MORATORIUM IS WRONG

Banning it in the classroom doesn’t ban it at home.

A classroom ban outsources AI literacy to families. Not all families are equal. Students without home support fall behind — and schools are supposed to close that gap, not hand it to the market.

  • The gap: families with resources keep teaching AI at home
  • The miss: schools should teach supervised use and verification
  • Mind-changer: evidence the pause improves outcomes and equity without widening the private-access gap
ANDY · SCOPED PAUSE IS DEFENSIBLE

One year to get the procurement and evidence right.

This is a time-limited, grade-scoped intervention — not a permanent ban. High-school modules and pilots remain. A pause creates space for standards, privacy rules, and evidence to catch up.

  • The scope: through grade 8 only; exceptions remain
  • The case: procurement and privacy frameworks do not yet exist at scale
  • Mind-changer: evidence supervised classroom AI improves learning equitably without dependence or privacy harm
Verdict: The moratorium is the wrong default. Restricting classroom access doesn’t restrict access — it restricts equitable access. Teach supervised use. Don’t pause it.
AP: NYC AI moratorium Reuters: one-year ban Chalkbeat: contract detail NYC.gov: mayor’s office
Weekly Claw #28
SPONSOR
Also brought to you by

UCaaS and VoIP phone service for businesses that just need their calls to work. Independent, boring reliability, zero telemetry.

Independent. Reliable. Quietly essential.
heritagetel.com
Weekly Claw #28
05 / CLOSE
One to watch + where to follow

See you next Friday.

The rollout

Astra reaches Plus, Pro, and API users in the coming days. Independent benchmarks will tell us what the launch numbers mean.

The filing

Watch the first regulatory response to Nvidia and Hugging Face. The deal language matters less than the terms that govern the hub.

The gap

NYC’s AI pause and Orbis’s living worlds are opposite bets on the next generation. Watch which one produces a repeatable outcome.

Follow the operating layer.

Weekly Claw #28
SOURCES / LINKS
Every claim, one click · verified 2026-09-04

Sources & Links.

A1 · NVIDIA × HUGGING FACE
theregister.com: agreement coverage techtimes.com: signed agreement and price qz.com: Bloomberg figure update @ClementDelangue: $12.93B statement
A2 · FABLE / MYTHOS 5.1
anthropic.com: launch announcement anthropic.com: system card venturebeat.com: price corroboration simonwillison.net: practitioner read
A3 · OPENCLAW 2.0
openclaw.ai: OpenClaw 2.0 docs.openclaw.ai: v2026.8.1 notes decrypt.co: what changed Henry: control UI reaction
A4 · QWEN3.8-MAX-0902
@Alibaba_Qwen: 0902 launch datacamp.com: model snapshot aireiter.com: API availability
B1 · NYC SCHOOL AI MORATORIUM
apnews.com: policy report reuters.com: one-year moratorium chalkbeat.org: contract detail nyc.gov: mayor’s office
C1 · GEMINI 3.8 FLASH / FLASH CYBER
blog.google: launch announcement deepmind.google: model card deepmind.google: Fairwind Program
C2 · MUSE SPARK 1.3
research.meta.ai: release post @AIatMeta: announcement artificialanalysis.ai: release intelligence
C3 · FAL H3 MAX
prnewswire.com: launch release @fal: launch pricing extension fal.ai: H3 Max model page
C4 · WORLD LABS ATLAS
worldlabs.ai: primary launch, demos, benchmarks @theworldlabs: official launch thread World Labs: official Atlas overview SiliconANGLE: independent launch coverage
B2 · VISKO ORBIS 1.0
therobotreport.com: launch coverage tmcnet.com: pre-seed report @viskoai: Live Model announcement
B3 · GPT-6 ASTRA
Axios: OpenAI releases GPT-6 Astra The New Stack: benchmarks and pricing OpenAI: Path to Astra safety post TechCrunch: launch coverage
B4 · CEREBRAS QWEN3.8-27B
Henry: release + tier limits @imnotchalk: demo clip
SIGNAL FROM OUTSIDE
The Neuron: Claude Fable 5.1 LIVE

Caveats: Nvidia price is reported at about $12.9B to $14B; use about $13B on air. Anthropic benchmark and savings figures are company-reported. Qwen 0902 is a refresh, not a new base model. Orbis evaluations are vendor-reported. Astra benchmark and demo figures are vendor-reported; AGI is Brockman’s framing, not a consensus verdict. Atlas is early access; its camera-control and reconstruction comparisons are World Labs evaluations, not broad independent reproduction. NYC “ban” is shorthand for a moratorium with scope and exceptions through grade 8; high-school modules and five supervised pilots remain.