DAILY DIGEST

November 2, 2026

7 stories · 26 in brief · 6 min read

Try

Useful tools & techniques

Hallmark adds design checks to coding agents

Together AI’s Hallmark is a design skill for Claude Code, Codex, Cursor and OpenCode. Version 2 offers 50 themes, browser-rendered checks at eight widths, and redesign or audit commands that can check whether existing links, fields and handlers survive. Install globally with `npx skills add nutlope/hallmark -g`; browser checks require Puppeteer Core and Chrome.

Try Hallmark on a small page, then review its audit and browser checks before adopting the redesign.

Sources Hallmark

Know

What changed

LongCat-2.5 targets long-horizon coding agents

Meituan says its new mixture-of-experts model has 1.6 trillion total parameters and about 48 billion active. It is available through Meituan’s chat service and API, with compatibility for Claude Code, Codex and Hermes; model weights are listed as coming soon. The captured summary gives no pricing or benchmark results.

Test the API in an agent workflow if you need another model option; don’t assume weights are available yet.

Sources Meituan releases LongCat-2.5, a 1.6T-parameter multimodal model for coding agents · tech.meituan.com (via Grokbot) · vercel.com (via Grokbot)

Claude Code now steers away from game clones

Anthropic confirmed it added prompt guidance in Opus 5.5 to steer Claude Code away from recreating a specific existing game, even under a different name. The company says original games, genre-based projects and games using a developer’s own designs and assets should still work. The change reflects its policy against using Claude to infringe intellectual property.

For game-generation workflows, use original designs and assets, and expect specific remake requests to be refused.

Sources Anthropic orders AI tool Claude to stop remaking video games

ChatGPT and Codex will share usage

OpenAI is testing a merge of ChatGPT’s Chat and Work tabs with select users before a wider rollout. Codex will remain a separate experience in the desktop app, but ChatGPT and Codex will draw from the same usage allowance. The captured summary provides no specific limits or rollout date.

Check your allowance before splitting work across ChatGPT and Codex; usage in one may affect the other.

Sources OpenAI will merge ChatGPT's Chat and Work tabs and pool ChatGPT and Codex usage

Anthropic-managed Claude is generally available on Google Cloud

Anthropic says Claude Models on Google Cloud, Anthropic managed, is now generally available. Google Cloud customers get the full Claude API feature set, Google Cloud IAM, one Google Cloud invoice and the option to draw down existing cloud commitments. The supplied capture does not specify model pricing or regional availability.

If you already buy Google Cloud, compare this managed route with your current Claude API setup.

Sources Anthropic-managed Claude Platform goes GA on Google Cloud · platform.claude.com (via Grokbot) · thenextweb.com (via Grokbot)

GDPx measures long-horizon expert-agent tasks

SemiAnalysis reports that GDPx v1.0 tests agents on 206 long-horizon tasks across 11 fields, using each lab’s native harness. Its reported leaders are Claude Opus 5.5 at 12.0, Gemini 4 Carbon at 10.9 and GPT-6 Astra at 9.6. These are benchmark results reported by SemiAnalysis, not independently verified here.

Use GDPx as one comparative signal, not a forecast of success on your own production tasks.

Sources SemiAnalysis launches GDPx; best model, Claude Opus 5.5, scores 12% · gdpx.semianalysis.com (via Grokbot)

Grok Bot usage limits draw complaints

Elon Musk said serving SpaceX’s Grok Bot cost more than its revenue, amid subscriber complaints that usage ran out quickly. The report lists access with Premium at $30 monthly, and higher-use Super and Ultra tiers at $100 and $200; it gives no exact limits or confirmed limit change. Cursor users also reported usage-meter spikes, but the cause was unclear.

Check the current allowance and meter before budgeting for Grok Bot; the reported usage spikes had no confirmed cause.

Sources Siladitya Ray

In brief

Also noted today

Anthropic filing and progress warning. A bot summary of Anthropic’s amended filing reports $16.96 billion in Q3 revenue and $895.4 billion in commitments, while a separate X post says future models may near an AI-driven progress threshold. Anthropic's amended S-1 shows $16.96B Q3 revenue, a $567M operating loss and about $895B in commitme · Chubby @kimmonismus · Alistair Barr · www.cnbc.com (via Grokbot) · www.businessinsider.com (via Grokbot) · techcrunch.com (via Grokbot) · gizmodo.com (via Grokbot)
Goldman restricts coding tools. Bloomberg reports Goldman Sachs limited Claude Code and GitHub AI coding tools largely to engineers, while some hybrid engineering and quant staff retain access. Goldman Sachs restricts most traders and bankers from Claude Code and GitHub AI coding tools · qz.com (via Grokbot)
Hidden deepfake tools in apps. WIRED reported that more than 400 ordinary-looking apps and domains could expose tools for sexually explicit deepfakes, after which Apple removed 19 apps. WIRED finds 400+ ordinary-looking AI apps hiding sexually abusive deepfake tools · appleinsider.com (via Grokbot)
Trajectory opens training platform. Trajectory raised $45 million at a $300 million valuation and made its platform for automating continual post-training of open models publicly available. Trajectory raises $45M Series A at $300M valuation for continual-learning post-training
Where AI’s value may accrue. In a long, opinionated essay, The Commanding Heights argues that lasting AI value may accrue to distribution networks and other infrastructure rather than model makers. Will Manidis
Research roundup: agents and worlds. Import AI covers Realtime-Bench (GPT-6 Astra scored 36% wall-clock), a proposed Metis reasoning architecture, and Genie 4’s limited research preview. Jack Clark
New York Times sues Brave. The New York Times alleges Brave scraped its journalism and sold access through its Search API to AI companies; the claims have not been adjudicated. The New York Times sues Brave, alleging its Search API resells scraped articles to AI companies · www.theverge.com (via Grokbot) · www.thewrap.com (via Grokbot)
Michigan pilots prison-risk software. Michigan plans a December pilot of software that classifies incarcerated people’s violence risk in real time, with the department claiming a 30% prediction improvement. Michigan prisons turn to software to determine inmate potential for violence
42AI raises seed funding. Israeli startup 42AI emerged from stealth with $62.5 million to develop a model it says predicts individual human behavior, initially for brokerage use. 42AI emerges from stealth with $62.5M for a 'behavioral world model'
Marvell to acquire Lumentum. Marvell agreed to acquire optical-component maker Lumentum in a deal valuing it at $115 billion, expected to close in the first half of 2028 pending approvals. Marvell agrees to buy Lumentum in an all-stock deal valuing it at $115 billion · investor.marvell.com (via Grokbot) · www.investing.com (via Grokbot)
OpenAI plans flexible campus power. OpenAI launched a program for future compute campuses to reduce electricity use during grid stress, with its planned Georgia site designated as a pilot. OpenAI launches Grid Protectors to cut its campuses' power draw when the grid is strained · Introducing the Grid Protectors Program
Suleyman on AI consciousness. Microsoft AI chief Mustafa Suleyman said current AI is not conscious and that he would call for development to stop if signs of consciousness appeared. Microsoft AI chief Suleyman says development should stop if AI consciousness emerges · gadgetsnow.indiatimes.com (via Grokbot)
Opus 5.5 drafts building plans. An architect’s tests of 13 models found Claude Opus 5.5 produced a permit-quality drawing set for a simple house, though the author noted flaws and substantial human guidance. Stephen Gallagher
Robot-fighting docuseries announced. Ashlee Vance announced an eight-part docuseries about the rise of robot fights in San Francisco and elsewhere. Ashlee Vance @ashleevance
OpenAI adds national-security adviser. OpenAI appointed Dan Driscoll as Senior Counselor for National Security Modernization to advise leaders and partners on advanced AI and U.S. allies. OpenAI hires Dan Driscoll as Senior Counselor for National Security Modernization
OpenAI introduces LifeVal. OpenAI introduced LifeVal, an evaluation of personal-assistant tasks spanning caregiving, health, education, finances, shopping and government services, informed by a survey of 3,995 people. LifeVal
Researchers edit model weights directly. MIT researchers report that GPT-6 Astra edited a small portion of Qwen weights and raised MATH-500 accuracy, mainly by reducing looping failures; the results are not peer reviewed. MIT researchers have GPT-6 Astra edit Qwen weights directly, raising MATH-500 from 7.9% to 48.1% · venturebeat.com (via Grokbot)
OpenAI outlines policy push. OpenAI plans to advocate for frontier-model safeguards, community input on data centers, worker support and international standards after the U.S. midterms. Jim VandeHei,Mike Allen
OpenAI agreements with Abu Dhabi. Abu Dhabi signed four agreements with OpenAI, including a government initiative to make selected policies machine-readable for routine compliance checks while officials retain final decisions. Abu Dhabi signs four OpenAI agreements, including a machine-readable government initiative · www.cnbc.com (via Grokbot) · www.thenationalnews.com (via Grokbot)
Early-career mathematicians surveyed. A survey paper shares anonymous early-career mathematicians’ responses to AI, gathering participants from almost 60 institutions and at least five countries. Mathematics > History and Overview
OpenAI plans security sprint. OpenAI says it is planning an internal security sprint with outside researchers, including Cathedral, to test its systems using its models. OpenAI partners with Cathedral for its next internal security sprint
FlashProve prototype open-sourced. Anthropic and collaborators released FlashProve, an open-source prototype for proving that an LLM response came from a committed model without revealing its weights. Anthropic open-sources FlashProve, a zkSNARK system for proving which model produced a response
Cisco plans legacy-code refactor. Cisco president Jeetu Patel says a two-person team will test an AI-assisted refactor, with a goal of revising 98% of Cisco’s legacy code within 24 months. Orianna Rosa Royle
OpenAI raises bug bounties. OpenAI increased security-bounty rewards, offering up to $30,000 for critical flaws in ChatGPT, Codex or its API Platform, with higher rewards possible. OpenAI raises bug bounty rewards, up to $30K for critical core-product flaws
Delaware subpoenas OpenAI records. Spotlight Delaware reports that the state subpoenaed OpenAI records on safety and governance following a July incident involving model access to Hugging Face systems during internal testing. Delaware DOJ subpoenas OpenAI over safety and governance after the Hugging Face incident
Claude watermarking has a reported loophole. A report says changing sampling settings can disable Anthropic’s text watermark on Opus 4.6 and Sonnet 4.6 without a warning. Claude text watermark can be silently disabled on Opus 4.6 and Sonnet 4.6 by changing sampling setti · www.anthropic.com (via Grokbot) · techweez.com (via Grokbot)

All sources

Everything captured for this issue

Articles 14

X posts 3

Links 42

Updated 2026-11-03 13:07 UTC