1 The news2 For your work3 Money4 Before you go$20 billion Image: CNBC OpenAI The ChatGPT maker is now seeking about $20 billion at a $1.25 trillion valuation, the FT reports, as investors grow wary of heavy AI infrastructure spending.
OpenAI has cut the target for its latest funding round by $10 billion to about $20 billion, the Financial Times reported on Friday, citing people familiar with the matter. Earlier reports said the company was raising $30 billion at a $1.4 trillion valuation. It is now seeking a valuation of $1.25 trillion. A consortium of UAE investors that includes the state-backed fund MGX is expected to be the largest participant, and OpenAI would gain an option to lease new computing capacity in the region. The Wall Street Journal reported that $11 billion had been secured so far and that some large investors had responded tepidly. OpenAI, which is pursuing a listing in 2027, had not commented when TechCrunch asked.
New target about $20 billion
Valuation sought $1.25 trillion
Secured so far (WSJ) $11 billion
Largest participant UAE consortium including MGX Why it matters Weaker investor appetite for the largest AI lab may shape pricing, capacity investment and the stability of a supplier many depend on.
OpenAI's funding round as reported Report Target Valuation Earlier report $30 billion $1.4 trillion Later report $30 billion $1.25 trillion Financial Times, Friday $20 billion $1.25 trillion
Widely reported TipRanks WOWTALE
Part 1 of 4 The news What changed in AI today, lab by lab, and who answers for it.
OpenAI says that from December, usage on Plus, Pro and Business plans will come from a single allowance shared by ChatGPT and Codex, replacing separate limits. The company says the change is part of its work to bring Work mode into Chat. Some users have asked to keep limits separate.
Plans affected Plus, Pro, Business
Starts December Why it matters Teams that use Codex heavily for coding could run through allowances meant for everyday ChatGPT work, so budgets need checking.
Axios reports that after the midterms, OpenAI will press Washington for mandatory safeguards on the most powerful AI systems, which it calls speed limits. Chris Lehane, its chief global affairs officer, said people should stay in control of the technology. OpenAI also wants data centres renamed AI compute campuses.
Why it matters If a leading developer backs binding limits, it could shape what a new Congress demands of every frontier model provider.
Shareholder Phil Rosen has filed a derivative lawsuit against Google chief executive Sundar Pichai and directors including Larry Page, Sergey Brin and John Doerr. Like his earlier suit against Apple's leadership, it centres on copyright. It also adds a claim under Illinois's biometric privacy law, BIPA.
Case Rosen v. Pichai
Claims Copyright, BIPA Why it matters Derivative suits test whether directors can be held personally responsible for their companies' AI data practices, adding to the cost of copyright disputes.
John Ternus, Apple's chief executive, told Wired the company does not want to ship AI for its own sake, and rejected criticism that it has fallen behind its rivals. He said Apple will use AI in new products, and that the test is whether customers really love what it builds.
Why it matters Apple's pace decides when AI features reach the iPhones and Macs that many organisations' staff already use every day.
Image: Tesla North Users of xAI's Grok Bot said their weekly usage allowances ran out days before the reset. Elon Musk replied that the cost of serving the bot had been unsustainably high relative to revenue, which suggests the faster drain was intended. He has not announced a reset or new limits.
Why it matters Providers are rationing flat-rate AI subscriptions as serving costs rise, so workflows that depend on them need fallback plans.
Image: The Guardian Other labs · National Grid Sky News reports that National Grid Electricity Transmission has agreed to buy more than 70 acres at Dorman Point, on the Teesworks site, for what would be Britain's largest substation. The new capacity is meant to help the north-east attract large AI infrastructure, including a planned datacentre nearby.
Land more than 70 acres, Dorman Point
Contractor in line Balfour Beatty Why it matters Grid connections are the main limit on building UK AI data centres, so new transmission capacity decides where compute can go.
03 Rules, safety and accountability Regulation, sovereignty, safety, ethics and liability.
A federal judge has denied xAI's motion to dismiss an NAACP lawsuit. The suit alleges the company ran at least 35 combustion turbines at its Colossus data centre near Memphis without the required air permits, in breach of the Clean Air Act. The Southern Environmental Law Center brought the case.
Claim Clean Air Act
Turbines alleged at least 35
Site Colossus, near Memphis Why it matters On-site gas power is a common shortcut for building data centres quickly, and this ruling raises the permitting risk of that approach.
Individuals and organisations asked Judge Stein this week for permission to file amicus briefs in the consolidated copyright litigation against OpenAI. The blog ChatGPT Is Eating the World, whose author is one of the applicants, counts 52 proposed briefs by its own unofficial tally, an unusually large number.
Proposed briefs 52 Why it matters So many outside briefs suggest the case could set widely cited rules on whether training models on copyrighted work is lawful.
Sovereignty · Government of India India's electronics ministry, MeitY, launched a handbook in Bengaluru on its Data Empowerment and Protection Architecture for AI training. Under the scheme, models are sent to where data is held instead of datasets being transferred. Officials say this protects privacy, commercial confidentiality and institutional control.
Issued by MeitY
Approach Model travels to the data Why it matters Organisations holding Indian data may be able to offer it for model training without handing the data itself to developers.
The Information reports that on Friday the White House Office of the National Cyber Director told the Protect Democracy Project it would not produce key documents on the administration's voluntary AI testing framework. This was despite a September agreement reached after a Freedom of Information Act lawsuit.
Why it matters Without the text, developers and buyers cannot see which tests frontier models face under the voluntary federal regime.
Risk · Western Australia Police A five-month trial of live facial recognition by Western Australia police scanned 1,380,882 faces at 43 locations, leading to 122 arrests and nine incorrect alerts. Police point to a very low error rate. Civil Liberties Australia says the scans were made without reasonable suspicion and the rate should be judged on convictions.
Faces scanned 1,380,882
Arrests 122
Incorrect alerts nine
Locations 43 Image: BritBrief Why it matters Public bodies elsewhere considering live facial recognition will face the same dispute over how accuracy and proportionality are measured.
Ahead of the US midterm elections, a study by Forum AI, a start-up that evaluates AI models, found that almost 40 percent of chatbot answers to election questions contained at least one factual error, AFP reports. Experts say chatbots can debunk conspiracies but often draw on unreliable web sources.
Answers with errors almost 40 percent Why it matters Organisations using chatbots to give the public information should expect errors on time-sensitive topics and should check the answers.
Wired reports that hikers who use chatbots to plan backcountry trips are being led astray, with potentially deadly results. The article argues that handing planning to AI can lead people to give up their own critical thinking in places where mistakes are hard to recover from.
Why it matters Relying on confident but wrong answers is a risk wherever staff use AI to plan work where safety is at stake.
Y Combinator and Olam Labs Y Combinator and Olam Labs gave 50 agents running OpenAI's GPT-6.1 Sol three hours in a Minecraft survival world to build a civilisation. After zombies killed many on the first night, the group turned to permits, rationing and procedure; four agents spent 11 minutes appointing a fertiliser operator.
Model GPT-6.1 Sol
Run 50 agents, three hours, one colony
Bed Wars GPT-6.1 Sol beat Claude Opus 5.5 six games to three
Status Investor lab write-up, not peer-reviewed Image: Tom's Hardware Why it matters Teams deploying multi-agent systems may see the same drift: agents given vague goals favour process over output, and coordination costs tokens.
Part 2 of 4 For your work Tools to use today, one thing to try, security to act on, and how work is changing.
Vercel Labs Open source Solo Team
Vercel Labs has released agent-desktop, a Rust command-line tool that lets coding agents read and operate desktop apps on macOS, Windows and Linux. It works through the operating system's accessibility tree rather than screenshots, returning JSON references for each element an agent can act on.
Use it Let a coding agent test or drive a native desktop app by acting on accessibility-tree references.
Licence Apache-2.0
Runs on macOS, Windows and Linux
Install npm install -g agent-desktop
Built with Rust Image: GitHub (vercel-labs) Why it matters Accessibility-tree control can be steadier than pixel matching, but it gives an agent real reach into the apps on a machine.
Presentations.AI has made a presentation-generation API and a white-label generator available, so software firms, agencies and agents can create editable PowerPoint decks inside their own products. The reference price is $0.10 per slide, halved to $0.05 for three months after launch; webhooks can trigger decks from CRM data.
Use it Generate a client deck automatically when a CRM opportunity changes stage, using the API's webhook triggers.
Price $0.10 per slide
Launch offer $0.05 per slide for three months
At launch offer $50 per 1,000 slides
Output Editable PowerPoint Why it matters Teams building sales or reporting tools can buy deck generation by the slide rather than building their own pipeline.
OpenAI's API changelog says eligible ChatGPT Pro and Business subscribers can claim up to $500 per month in free API credits by linking an API organisation they own. The same update makes the Decisions API generally available with gpt-6-luna, raising requests to 400 questions from 200.
Use it Link an API organisation you own to your ChatGPT plan, then spend the credits on prototypes.
Credit Up to $500 per month
Eligibility Eligible ChatGPT Pro and Business subscribers
Decisions API limit 400 questions per request
Previous limit 200 in public beta Why it matters Teams already paying for ChatGPT seats may be able to prototype on the API without a separate budget.
Vercel Labs Install agent-desktop and read its agent skill instructions These commands install Vercel Labs' open-source agent-desktop tool and print the skill file that explains its snapshot, reference and retry loop to an agent. Reading it first shows what the tool can do on your machine before an agent uses it; no account is needed.
Step 1
Install the tool globally with npm
npm install -g agent-desktopCopy Step 2
Check the installed version
agent-desktop versionCopy Step 3
Print the desktop skill: the loop, refs, verification and retry rules
agent-desktop skills get desktopCopy Wiz Research has named October 2026 'Black October' after at least nine AI developer platforms, including Fireworks AI, Railway, Warp and Northflank, disclosed security incidents. Attackers stole API keys and cloud credentials, checked them and resold metered model access, running up over $100k in enterprise usage.
Platforms affected At least nine
Downstream abuse Over $100k at list price
Latest disclosure Northflank, 30 October What to do Rotate AI and cloud keys held on any platform that disclosed an incident, and set spending caps and alerts on model accounts. Check logs for the credential-check IP addresses Wiz published.
Cisco Talos has described Volta-Autodoxer, a Windows infostealer that posts stolen data to OpenRouter and asks a model to profile each victim: where they live, their job and what their accounts are worth. It defaults to Claude Opus 4.6, falling back to Kimi K2 or Gemini 2.5 Flash.
Platform Windows
Analysis published 31 October, Cisco Talos
Route to models OpenRouter What to do Treat unexpected outbound calls from staff machines to AI model gateways as a possible sign of compromise. Reset passwords and sessions on any infected device.
Proton will re-interview some staff about how they use AI at work, with interviewers drawn from leaders outside their team, 404 Media reports; the results may affect who keeps their job. Proton says it is not making layoffs, is still hiring and wants to find who needs support.
Interview dates 4 to 11 November
Company position No layoffs; still hiring Why it matters Assessing staff on AI use is becoming a management tool; how it is run affects trust, retention and fairness.
10 How companies work with agents Agents at work Enterprise buyers trust their AI agents but mostly keep them on a short leash. G2, the software review site, looked at 3,222 enterprise reviews of AI agents across 231 products from 2025 and 2026. It found that 77% of reviewers trust their agent, but only 7% would give it total autonomy, and most want guardrails or human checks in place.
Set autonomy per task. Decide for each workflow whether the agent acts alone, acts within limits or waits for a person to approve.Write the guardrails down. List which systems, data and actions the agent may touch, and make anything outside that list need sign-off.Keep a human checkpoint. Have someone review the agent's output before it takes effect in high-risk work, and loosen this only as its record builds.Put limits on spend. Cost was a common complaint in reviews of agent builders and support agents, so set budgets and watch usage from the start.How much autonomy enterprise reviewers on G2 would give an AI agent
Autonomy with guardrails 60%
Limited autonomy 21%
Human approves every action 11%
Total autonomy 7%
Part 3 of 4 Before you go One long read, the rest of the day in a line each, and what is coming.
11 One long read Worth twenty minutes.
JiTTests Meta Engineering · testing Meta reports on running its Catching JiTTests in production. The system generates tests for each code change while it is under review, filters out weak catches, and asks engineers only to confirm whether a change in behaviour was intended.
Why it matters Engineering leads weighing AI-generated tests, or deciding where people should sit in an automated review pipeline, will get the most from it.
Jepsen found 96 issues in ParticleDB 1.1.4, a database for AI workloads, and questioned how far LLM-written documentation can be trusted. Jepsen Aramco has completed its purchase of a 49% stake in the Saudi AI company Humain, paying about $2.6 billion. People Matters Sema4.ai released version 3.0 of its enterprise agent platform, which moves repeated agent tasks out of the model and into code that can be reviewed. citybiz Parameter showed how a shared ChatGPT link could take over a user's account; OpenAI fixed the flaw on October 26. Parameter 3 Nov
Consultation closes · Online The European Commission is collecting views from rights holders, generative AI providers and others on how AI affects copyright protection.
16–20 Nov
Conference · Berlin A machine learning and generative AI conference for practitioners, held in Berlin and online.
25–26 Nov
Conference · London A two-day AI conference in London aimed at business and technical leaders.
26 Nov
Deadline · Online Google Cloud has deprecated Vertex AI Extensions and will shut the service down after this date, so teams using it need to migrate first.
30 Nov – 4 Dec
Conference · Las Vegas Amazon Web Services' annual cloud conference, where AWS typically announces new AI and infrastructure services.
1 Dec
Rules take effect · Online From this date, new Microsoft 365 Copilot Business subscriptions bought through CSP will have usage-based billing turned on by default; the date was moved from November 2.
