Episode 33 · Friday, October 9, 2026 · 4:00 PM ET / 20:00 UTC

Weekly Claw

Mistral previewed a 1T model; Reflection’s Beam is 501B, and both weights remain pending. OpenAI’s public math corpus has a correction trail. Windows and workplace agents are expanding. The debate: can open agents feel mainstream without giving up user control?

Hosts: @AndyML · @HiM ~37 min · three grids · one anchor · one debate

Sponsored by Herald Labs Heritage Telecom

“Volume without verification.”

Weekly Claw #33

01 / COLD OPEN

Cold open · the frame

The week the receipts went public.

Large model previews still have weights pending; a public math corpus shows withdrawals and repairs. Windows and Google are expanding agent workflows. Henry and Teknium disagree on a consumer retreat: can open harnesses match mainstream ease while preserving user control? Benchmark, policy, and security reports round out the week.

1

Weights pending

2

Math with revisions

3

Agents enter Windows

4

Workplace agents

5

Measure and route

6

Policy, security, debate

🧠Mistral previewed 1T total / 52B active; Reflection’s Beam is 501B / 23B active. Both sets of weights remain pending.

💥OpenAI’s repo listed 722 manuscripts; three withdrawals left 719, and 14 others were repaired. No peer review is claimed.

📜Windows 11 gets GA execution containers for OpenClaw; Google announced a shared workplace agent with persistent cloud execution.

🛡Henry reads the polished UX of Grok Bot, Muse, and Dot as pulling consumers from open harnesses; Teknium disputes that reading.

Weekly Claw #33

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.

heraldlab.si

Weekly Claw #33

WHAT HAPPENED THIS WEEK

What happened this week · part 1

Open weights, public math, Windows for agents.

Mistral and Reflection previewed large models with weights pending; Microsoft adds Windows infrastructure for agents. OpenAI’s math corpus and a separate 3SUM paper carry distinct records.

Artificial Analysis page for Mistral Large 4 Preview: intelligence index 38, $1.36 in and $4.18 out per million tokens, 'notably fast, however very verbose'

Mistral previews Large 4: 1T total, 52B active.

Weights are promised for the end of October and remain unreleased. Artificial Analysis lists an index of 38 and API prices; those are separate from Mistral’s own preview claims.

▶ PRIMARY · MISTRAL LARGE 4

Reflection's own Beam chart: 501B parameters, 23B active, scored against mean generated tokens on DeepSWE, HLE text-only, Terminal Bench 2.1 vs GLM-5.2 and Qwen3.8

Reflection previews Beam: 501B total, 23B active.

Weights, technical report, and model card are promised later this month. The coding and reasoning chart is Reflection’s own; independent runs await released weights.

▶ PRIMARY · REFLECTION BEAM

$GitHub repository openai/math: commits updating manuscripts and Lean formalizations, Apache-2.0 license$

OpenAI’s math release: revisions and authorship.

722 manuscripts across 372 families; three withdrawn by Oct 7 left 719, and 14 others were repaired. 300/719 (~42%) list top-line formalizations. In the separate 3SUM paper, Claude discovered the algorithm; Josh Alman and Virginia Vassilevska Williams developed and authored the paper.

▶ REPO HISTORY FORMALIZATION OVERVIEW 3SUM PAPER

Windows
Hybrid intelligenceExecution Containers: GA
HydraFusion: experimental preview later October

Microsoft adds agent execution and routing to Windows.

Execution Containers are GA on Windows 11; a native Windows gateway supports OpenClaw. HydraFusion local/cloud routing is experimental preview later in October; Copilot+ PC context, actions, and models are planned for coming months.

▶ PRIMARY · BUILDING WINDOWS FOR HYBRID INTELLIGENCE

Weekly Claw #33

WHAT HAPPENED THIS WEEK

What happened this week · part 2

Measure outcomes, route work, give agents a workplace.

One scoped benchmark, native X access, and workplace-agent announcements put new agent workflows to the test.

Hugging Face article on ThinkingBox: in a common-set ablation covering 121,680 valid trials across 12 LLM models, 79,853 attempts failed

ThinkingBox reports failures in a common-set ablation.

79,853 of 121,680 trials failed across 12 models (~65.6%) in the paper’s common-set ablation. Treat this as that measured setup, not a universal agent failure rate.

▶ PRIMARY BLOG · PAPER 2608.19741

Grok BotNative X search, reading and monitoring
Selected model routing

Grok Bot adds native X access and selected model routing.

Native X search, reading and monitoring for all users, without configuring the X connector. Cursor lists Opus 5.5 routing; tasks may route to different models. No per-user picker or added per-token charge merely for routing.

▶ X ANNOUNCEMENT FOLLOW-UP CURSOR DOCS

Gemini at Work 2026Shared workplace agent
Early access announced

Google introduces the Gemini workplace agent.

Gemini at Work introduces persistent cloud execution, coworker agents with shared memory and dedicated identities, Gemini/Claude orchestration, and spend caps with pause controls. SMB early access starts now; broader access is planned soon. No universal price or date is given.

▶ PRIMARY · GEMINI AT WORK 2026

Nous Research$90M fundraise
Hermes for Businesses: intent

Nous confirms a $90M fundraise; its business product is still a plan.

Nous Research develops Hermes. Its note signals intent to bring Hermes to businesses; it does not announce a launched product.

▶ PRIMARY · NOUS FUNDRAISE NOTE

Weekly Claw #33

WHAT HAPPENED THIS WEEK

What happened this week · part 3

A policy update, a cache-layer RCE, and a .claw application.

Three distinct status updates: Anthropic revises its usage policy, a network-exposed LMCache transport has a critical RCE advisory, and ICANN lists an active application for the .claw top-level domain.

Anthropic's 2026 Usage Policy update page, Oct 8 2026: the paragraph adding a prohibition on sustained and needless abusive or cruel behavior toward our models

Anthropic updates its usage policy on model-directed abuse.

Announced Oct 8, effective Nov 12. The primary announcement says enforcement centers on ending conversations; carve-outs include ordinary frustration, creative themes, and research.

▶ PRIMARY · POLICY UPDATE

JFrog Security Research advisory JFSA-2026-001694382: LMCache vulnerable to unauthenticated remote code execution via pickle deserialization on the multiprocess ZMQ transport, CVE-2026-105192, CVSS 9.8, published Oct 7 2026

LMCache has a critical RCE in its multiprocess transport.

An unauthenticated pickle message can run code with the process’s privileges when the ZeroMQ service is network-exposed. Official containers run as root; no fix was listed as of Oct 7. Default localhost binds are not remotely reachable.

▶ JFROG ADVISORY · CVE-2026-105192

@steipete · Oct 8 · 18:50“WE GOT IT! .claw incoming!”Quoting @drodecker’s winning-applicant report

Peter announces “.claw incoming.”

He quoted David Rodecker’s report naming the OpenClaw Foundation as winning applicant. ICANN still lists Active — Pre-Evaluation Processing; final delegation and domain-registration availability are not established.

▶ PETER’S POST DAVID’S REPORT ICANN STATUS

Weekly Claw #33

03 / SIGNAL FROM OUTSIDE · 6 MIN

Signal From Outside · weekly video review · permanent anchor · 6 min

The practitioner canary: “I love Ultrafast (it’s unusable).”

THIS WEEK’S SIGNAL · THEO · T3.GG

“I love Ultrafast (it’s unusable)” — the speed tier priced honestly.

Theo edits an Astra app in real time, then looks at the bill: $50/M output tokens becomes $300/M on Ultrafast and $450/M on long context. His reaction: “These bills are actually insane.”

Those prices and the workflow are one practitioner’s report, not a universal benchmark. The clip ties speed to cost in a concrete development task.

▶ WATCH ON YOUTUBE · 28:17

THEO · T3.GG · QUOTED OUTPUT-TOKEN PRICES / MILLION

“I love Ultrafast (it’s unusable).”

$50/MAstra

$300/MUltrafast

$450/MLong context

Price ladder and workflow from Theo’s video; one practitioner’s workload.

Theo discusses an Astra edit alongside the quoted price ladder.

Weekly Claw #33

04 / HOT TAKE · 4 MIN

Hot take · two views, one question

Who wins the consumer agent?

HENRY · OCT 7 THESIS

Polished UX is pulling consumers toward closed agents.

Henry argues that users are leaving open-source harnesses for Grok Bot, Muse, and Dot. He reads OpenClaw’s enterprise emphasis and Hermes’ coding/science focus as a retreat from consumers. That is his interpretation, not measured market-share data.

▶ HENRY’S POST · OCT 7

TEKNIUM · OCT 6 POSITION

Wider-use focus, with consumer mobile plans.

Teknium says consumers are not the only demographic: scientists, cybersecurity, developers, and knowledge workers matter too. He says mobile will be almost exclusively consumer-focused.

▶ TEKNIUM’S WIDER-USE POSITION · OCT 6