1 / 10
← → navigate · scroll / swipe · F fullscreen
Episode 35 · Friday, October 23, 2026 · 4:00 PM ET

Weekly Claw

Open models beat GPT-5.6 Sol on agentic coding this week — and Kimi matched Opus-class terminal work — then Washington removed DeepSeek from every civilian federal system and Google Cloud retired all sixteen of its hosted open-model endpoints on the same day it added OpenAI. Anthropic’s draft S-1 showed a $750B compute bill — then it pledged 5% of its shares to public good and Musk answered with a SpaceX share for every American kid. Meanwhile one hacker’s agents broke into 100+ companies for about $50 a breach, a bad LiteLLM wheel nearly took the keyring with it, and php2shell put unauthenticated RCE into PHP core — patched the same week.

Hosts: @AndyML · @HiM ~31 min · three grids · one debate
Sponsored by Herald Labs Heritage Telecom
“Open weights won the benchmark. Then they lost the building.”
Weekly Claw #35
01 / COLD OPEN
Cold open · the frame

The open lane got the best model of the week. Then it lost every comfortable place to run it.

A 42.3 on Terminal-Bench 4.0 — above GPT-5.6 Sol — shipped under MIT on a Saturday morning, and on show day itself Kimi’s 2.8T-parameter K3.1 Pro posted 65.7 on the same bench — between GPT-6 Astra and Opus. Seven days later that same lab is banned from federal IT, its search API is a sixth of the incumbent price, and the hyperscaler everyone expected to keep open models alive retired all sixteen of them. The closed labs published the reason in a leaked prospectus: a $750B compute bill.

1
Open models win the bench
2
The eviction notices
3
The compute bill
4
Agents at war
5
One to watch
📈DeepSeek shipped V4.1-Flash-1017 under MIT: Terminal-Bench 4.0 at 42.3, above GPT-5.6 Sol’s 39.9. Opus 5 is still ahead at 51.8.
🏛DHS ordered DeepSeek removed from all civilian federal IT — the first government-wide exclusion of a Chinese frontier lab.
💰Anthropic’s draft S-1: $16.3B in H1 revenue, an $80.7B net loss mostly on paper, and up to $750B in future compute commitments.
🔧Amazon open-sourced TorchNeuron, Google open-sourced TorchTPU, and Tenstorrent raised $1.5B at $6.2B — three layers of CUDA’s moat attacked in one week.
🛡One operator’s agents broke into 100+ companies for roughly $50 a breach, while a compromised LiteLLM wheel on PyPI went after the keys that run everyone’s fleet.
Weekly Claw #35
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.
heraldlabs.si
Weekly Claw #35
WHAT HAPPENED THIS WEEK · PART 1
What happened this week · part 1

Open models won the benchmark and lost the catalogue.

Three open launches cleared the frontier bar this week. Four days later the federal register and Google Cloud each published a removal. Each card links its primary receipt.

DeepSeek-V4.1-Flash-1017 model-card benchmark table: Terminal-Bench 4.0 42.3, DeepSWE v1.1 77.0, ProgramBench 29.1, HLE 42.6, Agents' Last Exam 35.4, against V4.1-Flash, K3, GLM-5.3, GPT-5.6 Sol and Opus 5

DeepSeek ships V4.1-Flash-1017: an open model that beats GPT-5.6 Sol on Terminal-Bench.

MIT-licensed, 1M context, repo live 2026-10-17T02:40Z. Terminal-Bench 4.0 at 42.3 vs Sol 39.9 and GLM-5.3 37.9; Opus 5 still ahead at 51.8. The same weekend it opened a Search & Fetch API at $0.75 per 1,000 calls, free inside Claude Code and Codex until Nov 1. Thursday, WSJ: DeepSeek closed its second round at ~$15B — nearly double target — with a Shanghai STAR-listing prep under way.

▶ LIVE · huggingface.co · VENDOR BENCH, MIT WEIGHTS
Kimi tech blog, 2026-10-23: 'Kimi K3.1 Pro Tech Blog: Advancing Open Intelligence', with the 1M-token context and 2.8T/104B parameter lines visible

Kimi K3.1 Pro: a 2.8T MoE that lands between Astra and Opus on terminal work.

Posted 1:15 PM ET on show day. Terminal-Bench 4.0 at 65.7 vs GPT-6 Astra’s 58.2 and Opus 5.5 at 66.4; 1M context, multimodal, Blender/CAD and finance claims. Weights promised by Nov 4; Highspeed variant runs 2–3× faster.

▶ LIVE · kimi.ai · VENDOR BENCH, WEIGHTS NOV 4
Federal Register notice 2026-21447: 'Notice of Order Regarding Exclusion and Removal of DeepSeek Products From Federal Civilian Executive Branch Information Technology Systems and Procurements', Homeland Security Department, 10/20/2026

DHS orders DeepSeek removed from every civilian federal system.

Federal Register notice 2026-21447, published Oct 20: exclusion and removal across FCEB IT and procurement. Vendors now have to prove no transitive DeepSeek routing. Reuters and The Verge reported it a day earlier.

▶ LIVE · federalregister.gov · PRIMARY ORDER
Google Cloud documentation page 'Open model deprecations' for Gemini Enterprise Agent Platform, listing the retirement of managed open models on Vertex

Google Cloud adds OpenAI’s models and retires all 16 open-model endpoints the same day.

Oct 21: GPT-6 Astra, GPT-6.1 Sol and Luna land in Model Garden private preview. The same day’s deprecation notice retires every managed open-model endpoint — DeepSeek, Qwen, Kimi, GLM, Llama, MiniMax, gpt-oss. Open weights lose their easiest enterprise home.

▶ LIVE · cloud.google.com · 3 IN / 16 OUT
Weekly Claw #35
WHAT HAPPENED THIS WEEK · PART 2
What happened this week · part 2

The compute bill and the giving.

The leaked prospectus priced the closed-model era; days later the labs started pledging the winnings back. Each card links its primary receipt.

Yahoo Finance syndication of the WSJ story 'Anthropic revenue hits $16.3 billion as preliminary S-1 outlines IPO risks', with the article headline and deck visible

Anthropic’s draft S-1: $16.3B revenue, $80.7B loss, and up to $750B in compute commitments.

WSJ review of the draft prospectus, Oct 20: H1 revenue $16.3B from $1.1B, net loss $80.7B of which ~$72B is non-cash convertible-note accounting, ~$9B of real H1 infra spend, and up to $750B of possible commitments including non-cancellable $111B Google, $110B AWS, $31B Microsoft. Two customers are ~20% of revenue. Investors weighing ~$1.5T against ~$2T.

▶ LIVE · WSJ + Reuters review · DRAFT DOC, NOT FILED
Anthropic announcement page, Oct 18 2026: 'Investing in the public benefit' — 5% of shares committed to the Anthropic Public Benefit Fund, plus 2% for every 50% valuation step-up

AI giving: Anthropic commits 5% of its shares; Musk pledges a SpaceX share to every American kid.

Anthropic’s Public Benefit Fund (Oct 18): 5% of shares, +2% per 50% valuation step-up — ~20% of the company for public good all-in. Days later Musk pledged one of his 4.8B SpaceX shares to every US child, ~$11B. Weeks before IPOs, giving became competitive.

▶ LIVE · anthropic + wapo · PRIMARY PLEDGES
Weekly Claw #35
WHAT HAPPENED THIS WEEK · PART 3
What happened this week · part 3

Agents went to war — and the keyring got hit.

A week of agentic operations and AI-assisted break-ins: an attacker’s unit economics, a compromised gateway package, and php2shell in the language layer. Each card links its primary receipt.

NorForge blog 'Inside Cairn: Open Windows into massive AI-Driven Cyber Operations', Oct 19 2026, with the summary line about 20+ exposed Cairn panels

One hacker’s agents broke into 100+ companies for about $50 a breach.

Reuters and NorForge, Oct 19: a Chinese-speaking operator pointed Cairn agents at 1,200+ sites from Sep 25; 100+ breached, including Deere, SharkNinja and Cummins. Roughly 23 hours of unsupervised probing per success, a few thousand dollars of tokens. The Army is buying the same thing: ARCYBER’s ECHOTRIBBLE, still whitepaper-phase.

▶ LIVE · reuters + intercept · PRIMARY REPORT
GitHub issue 47826 in BerriAI/litellm: '[SECURITY] litellm 1.107.0 compromised (PyPI removed; Docker latest/main-stable removed): cook.pth credential stealer + sitecustomize RCE loader, built from lookalike repo'

A bad LiteLLM wheel on PyPI went after the keyring — and Railway froze every deploy.

Attackers stole GitHub CLI OAuth tokens from three LiteLLM maintainers, built 1.107.0 from a lookalike repo, and shipped a credential stealer plus an hourly backdoor. Caught and pulled the same day. Railway disclosed the same week: encrypted env vars accessed, some decrypted, all deployments paused, upstream AI keys revoked — “a series of attacks targeting platform companies to compromise AI inference providers.”

▶ LIVE · github + railway · PRIMARY DISCLOSURES
CERT-EU security advisory 2026-018: Critical Vulnerabilities in PHP, CVE-2026-109618 and CVE-2026-109619, both CVSS 9.2, patched 21 October 2026

php2shell: unauthenticated RCE in PHP core — found with AI, patched in days.

CVE-2026-109618/109619, CVSS 9.2, hit ~every PHP install serving requests — WordPress to Wikipedia — patched Oct 21 per CISA/CERT-EU. Node.js shipped high-severity fixes the same day, including a Brotli RCE. Decade-old flaws, unlocked by frontier models.

▶ LIVE · cert-eu + cisa · PATCHED, PoCs CIRCULATING
Weekly Claw #35
04 / HOT TAKE · 3 MIN
Hot take · the debate

Taste is the last moat. Everything else got automated.

Not a launch, not a breach, not a valuation: a claim about what is actually scarce now that a 2.5B-active open model writes production code and a $0.15-per-million API closes tickets.

Henry

If code is cheap, the scarce input is knowing what to build and what to throw away. The people who win the next two years are the ones with judgment about the product, not the ones with the fastest fingers. That is why the model race stopped being the interesting question.

Andy

Taste without distribution is a hobby. The moat is the boring stuff: who holds the keys, who carries the pager, who can prove the thing works. This week’s $50-a-breach story is a taste problem only if you have never been paged at 3am.

What would change Henry’s mind: one documented case where a taste-led, non-engineer team shipped and kept a production system for a year.
Weekly Claw #35
SPONSOR
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 #35
05 / CLOSE
One to watch + where to follow

See you next Friday.

The filing

Anthropic’s prospectus is expected to flip public around October 25. Everything aired tonight is from a draft reviewed by wires, not a filed document. Watch whether the $750B commitment ceiling survives contact with the public version.

The raise

DeepSeek closed its second round at ~$15B — nearly double target — with Tencent and CATL in, per WSJ; a Shanghai STAR-market listing is being prepped for as soon as early 2027. The lab Washington removed from federal IT may price its IPO anyway.

The free window

DeepSeek’s Search & Fetch calls are free inside Claude Code, Codex and the DeepSeek Harness until November 1. After that the meter starts at $0.75 per 1,000 — check your harness call volume before then.

Open weights won the benchmark. Then they lost the building.

Weekly Claw #35
SOURCES / LINKS
Every claim, one click · verified 2026-10-23

Sources & Links.

A1 · DEEPSEEK V4.1-FLASH-1017
huggingface.co: MIT weights, repo created 2026-10-17T02:40Z, vendor benchmark table api-docs.deepseek.com: Search & Fetch API guide, $0.75/1k, free window to Nov 1
A2 · KIMI K3.1 PRO
kimi.ai: K3.1 Pro tech blog, TB4.0 65.7 vs Astra 58.2 / Opus 5.5 66.4, weights by Nov 4 x.com: launch post, 2026-10-23 17:15 UTC
A2b · MINIMAX (WATCH ONLY)
minimax.io: Space Bunny reveal, DeepSWE 73.7, weights later huggingface.co: M3.1 Flash weights out Oct 23 — 428B/23B MoE, 252GB
A3 · DHS DEEPSEEK ORDER
federalregister.gov: notice 2026-21447, exclusion and removal, 10/20/2026 reuters.com: ban on DeepSeek in agency IT systems
A4 · GOOGLE CLOUD SWAP
cloud.google.com: OpenAI models in Model Garden, Oct 21 docs.cloud.google.com: open-model deprecations, all managed endpoints
B1 · ANTHROPIC DRAFT S-1
wsj.com: $16.3B revenue, $80.7B loss, IPO risks (paywalled; body via syndication) Reuters via Miami Herald: $80.7B loss, ~$1.5T valuation talk, November IPO
B2 · AI GIVING
anthropic.com: Public Benefit Fund, 5% + 2% per 50% step-up, Oct 18 washingtonpost.com: Musk + rivals pledge ~$300B total
B2b · CUDA ALTERNATIVES
aboutamazon.com: TorchNeuron, Apache-2.0, one-line Trainium port tenstorrent.com: $1.5B Series E at $6.2B Reuters via Yahoo: 3M+ RTX PRO 5500 orders, $35B+
C1 · AGENT WAR
norforge.com: Inside Cairn, 20+ exposed panels, Oct 19 theintercept.com: ARCYBER ECHOTRIBBLE autonomous agents
C2 · SUPPLY CHAIN
github.com: litellm 1.107.0 compromised, PyPI + Docker removed blog.railway.com: env vars accessed, deployments paused, AI-key motive nodejs.org: Oct 2026 security releases, Brotli RCE CVE-2026-111497
C3 · LANGUAGE-LAYER CVEs
php.net: 2026 archives — php2shell fixes in 8.5.12 / 8.4.27 / 8.3.36 / 8.2.35

Vendor-reported: DeepSeek + Kimi + MiniMax benchmarks, Nvidia fill rate; DeepSeek $15B raise per WSJ-sourced reports. Talks/draft-stage: Anthropic S-1 and giving pledges.