1 The news2 For your work3 Money4 Before you go$895.4 billion Anthropic Anthropic's revised prospectus adds third-quarter results: revenue of about $17 billion, slower quarterly growth, a return to operating loss and $895.4 billion in committed spending.
Anthropic filed an amended Form S-1 with the US Securities and Exchange Commission on Monday, adding the third-quarter results that its first prospectus, published last week, left out. Revenue for the quarter was about $17 billion, up about 1,127% on a year earlier, but growth on the previous quarter slowed to about 47%, from 144% in the second quarter. Total operating expenses rose to $17.5 billion from $10.6 billion, returning the company to an operating loss as computing costs climbed. Purchase obligations and leases now add up to about $895.4 billion. The filing also says Anthropic is considering updates to its Responsible Scaling Policy to keep humans in control as models approach recursive self-improvement, a threshold it says none has yet crossed.
Q3 revenue About $17 billion
Nine-month revenue $33.24 billion
Nine-month net loss $104 billion
Committed spending About $895.4 billion Why it matters Organisations building on Claude get a detailed view of the costs, commitments and risks behind a supplier they depend on.
Anthropic revenue growth on the previous quarter, 2026 Quarter Growth on prior quarter First quarter 132% Second quarter 144% Third quarter 47%
Widely reported CNBC Business Insider (via Yahoo Finance) gHacks Crypto Briefing
Part 1 of 4 The news What changed in AI today, lab by lab, and who answers for it.
OpenAI has launched a Grid Protectors Program under which its future US compute campuses would cut electricity demand when the grid is strained. The pilot is a planned campus in Effingham County, Georgia, where it pledges cuts of up to a gigawatt and says it will pay full infrastructure costs.
Pilot site Effingham County, Georgia
Demand cut Up to a gigawatt
Equivalent Roughly 700,000 homes Why it matters Opposition to data centres over power bills is growing; flexible demand commitments may become a condition of planning and grid approval.
OpenAI is folding ChatGPT's separate Work tab into the main chat, so users can ask it to build websites, make presentations or act on the web in one place. Select users get it first. ChatGPT and Codex will then share one usage allowance, so some may hit limits sooner.
Availability Select users first, then wider rollout
Codex Remains separate in the desktop app Why it matters Teams using both ChatGPT and Codex should expect a shared allowance and may need to plan around more frequent limit resets.
Anthropic has added two sentences to Claude Code's instructions telling it to build original or genre games, not recreations of a specific existing game, even renamed. A staff member said cloning breaches its usage policy. Developers had complained of AI-made copies flooding Steam; critics doubt one company can stem it.
Model Claude Opus 5.5 in Claude Code
Allowed Original games, genres, user's own designs Why it matters Shows model makers using system prompts to enforce intellectual property rules, which can also block legitimate work that resembles existing products.
Anthropic's amended IPO filing says the value of its own staff's Claude use, priced at customer rates, first exceeded monthly personnel costs, excluding stock-based pay, in April 2026. It says that internal usage grew more than 250 times between January and September, with engineers increasingly directing and reviewing AI-written work.
Crossed payroll April 2026
Basis Customer list prices; excludes stock-based pay Why it matters A rare disclosure of how heavily an AI developer uses its own tools, a benchmark for firms weighing token spend against staffing.
Google has launched Dreambeans, an experimental app that turns users' activity across its services into personalised visual stories, in India, its first market outside the US. It supports 13 Indian languages alongside English on Android and iOS, though each story can use only one language at a time.
Languages 13 Indian languages plus English
Platforms Android and iOS
Market First outside the US Why it matters India is Google's test of whether personal AI products built on Gmail and Photos data travel beyond English-speaking markets.
Oracle plans to raise $20-billion through bond and share sales to fund its AI and cloud build-out. Its shares fell 4 per cent before trading opened. It had US$125.34-billion in borrowings at 31 August and has pledged to raise no more than $20-billion in additional AI financing through fiscal 2030.
Raise $20-billion in bonds and shares
Borrowings US$125.34-billion at 31 August
Share move Down 4 per cent premarket Image: The Globe and Mail Why it matters Oracle's borrowing to build AI capacity is a gauge of how cloud suppliers fund the compute that enterprise customers rent.
Image: TechCrunch Elon Musk has confirmed that Grok Bot usage limits were tightened because running costs had become unsustainably high, after subscribers complained of exhausting allowances early and waiting long for resets.
Launched August
Paid weekly active users Over 800,000 Why it matters Agent subscriptions priced at flat rates are proving costly to serve, so users of any provider should expect limits to change.
Image: TechCrunch On Monday's earnings call Apple's finance chief, Kevan Parekh, said AI could lift revenue through device upgrades and through opportunities across third-party apps and partnerships, without giving detail. Investors are asking whether Apple might distribute third-party AI agents through the App Store and take a share of their revenue.
Why it matters If Apple takes a cut of AI agents sold through the App Store, developers' distribution costs and terms could change.
03 Rules, safety and accountability Regulation, sovereignty, safety, ethics and liability.
Liability · Brave Software The New York Times, New York Daily News, Chicago Tribune and Orlando Sentinel have sued Brave Software in Manhattan federal court. They allege Brave scraped their journalism into a search index sold to AI companies including Anthropic and Perplexity, and let users bypass paywalls. Brave did not immediately comment.
Index size alleged More than 30 billion web pages
Complaint 69 pages
Remedy sought Injunction and damages Why it matters The case targets a supplier of search data to AI firms, so companies buying such data may face questions about its provenance.
Judge Lewis Kaplan has refused to dismiss Ace Cam's claim that Runway AI broke the DMCA's anti-circumvention rule by bypassing YouTube's access controls to obtain videos. The ruling follows similar decisions by other federal judges, and the claim against Runway will now go forward in the New York court.
Case Ace Cam v. Runway AI
Claim DMCA 1201 anti-circumvention Why it matters Courts are increasingly letting creators pursue access-control claims over scraped video, a route separate from fair-use arguments over training.
Image: The Register Sovereignty · Government of India India's electronics ministry has proposed a Data Empowerment and Protection Architecture for AI training, under which hospitals, banks and agencies would place anonymised, encrypted data on exchange platforms where developers can train models without taking the data away. The government wants more models built on local data.
Proposer Ministry of Electronics and Information Technology
Status Proposal Why it matters Developers seeking Indian health, banking or public data may have to train inside government-approved platforms rather than copy datasets.
Image: The Guardian Western Australia police ended a five-month trial of live facial recognition vans that scanned 1,380,882 faces across 100 deployments at 43 locations. Police report 122 arrests and nine incorrect alerts, and officers engaged with 146 registered sex offenders. The trial concluded on Saturday, police said.
Faces scanned 1,380,882
Arrests 122
Incorrect alerts Nine
Duration Five months Why it matters A measured record of arrests and false alerts gives both supporters and critics of live facial recognition evidence to argue from.
Risk · Chinese-language app developers A Wired investigation found a network of more than 400 apps and websites, linked to Chinese-language developers, that present as ordinary AI tools but can unlock hidden features for generating sexually violent deepfakes from uploaded photos. Some passed review by Apple, Google and Meta despite policies banning such content.
Image: Wired Why it matters App store review missed concealed features, so organisations cannot assume vetted apps are safe for staff or customers to use.
Reward · Stony Brook University A Stony Brook University study found physicians preferred AI-generated hospital discharge summaries to human-written ones in 95 percent of evaluated hospitalisations, rating them higher for quality, readability, accuracy and completeness. Primary care doctors found them easier to use, though the researchers say both kinds of summary contained errors.
Why it matters Discharge notes are a heavy documentation burden; hospitals weighing AI drafting gain evidence, along with a reminder that review remains necessary.
Four training projects in Barnsley will share over £800,000 to bring practical AI skills into everyday workplaces, the UK government says. The money is aimed at groups outside the technology sector, including care workers, women returning to employment and builders, rather than specialist roles alone.
Projects Four
Funding Over £800,000 Why it matters Public money is going to AI skills for non-technical staff, a model employers in care and construction can draw on.
OpenAI has published LifeVal, an evaluation of how well AI helps with everyday tasks such as finding childcare, managing money and government paperwork. Tasks were chosen from large time-use studies and surveys of where people want help. OpenAI says it has used it for a year to drive model improvements.
Published 2 November 2026 Why it matters Gives buyers a test closer to real errands and paperwork than academic benchmarks, though it is built and scored by a vendor.
A Microsoft study published in Science reports that its diagnostic AI system correctly identified 83% of 304 complex New England Journal of Medicine cases, against 36% for generalist physicians. The test was a simulation, and reporting notes the results need closer examination before clinical conclusions are drawn.
Cases 304 from the New England Journal of Medicine
AI accuracy 83%
Generalist doctors 36% Why it matters Strong simulated results strengthen the case for AI diagnostic support, but clinical trials with real patients are still needed.
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 has added LongCat 2.5, Meituan's multimodal model for coding and multi-step tool use, to its AI Gateway. It takes text and images with a 1M-token context window, and is called as meituan/longcat-2.5 from the AI SDK or from coding agents after a setup command.
Use it Point an existing AI SDK call or coding agent at meituan/longcat-2.5 and compare it against your current model.
Context window 1M tokens
Inputs Text and images
Pricing Provider pricing, no markup or platform fee
Model name meituan/longcat-2.5 Image: Vercel Why it matters Teams comparing models can trial a further Chinese-built coding model without opening another provider account or contract.
Cloudflare has pushed the start of AI Search billing back one month, to December 1, 2026, with pricing and included usage unchanged. New instances now re-sync their sources every 24 hours by default instead of every 6 hours, though any interval from 15 minutes is still selectable.
Use it Index a docs site or bucket for retrieval, then set sync_interval in the API to match how often content changes.
Billing starts December 1, 2026
Included each month 5 million ingestion tokens, 1,000 semantic queries, 1,000 full-text queries
Default sync 24 hours (was 6 hours)
Sync range 15 minutes to 24 hours Why it matters Teams relying on fresher search results should set the sync interval explicitly before charges begin in December.
Amazon Connect Customer adds an External Tool block that lets contact-centre flows invoke Model Context Protocol tools from integrated applications through an Amazon Bedrock AgentCore gateway. Calls can run synchronously or asynchronously, with a second block loading the result, and success and error branches.
Use it Look up an order status through an MCP tool mid-call and route the customer by the result.
Requires An Amazon Connect Customer instance
Integration External application integrated as an MCP server
Modes Synchronous or asynchronous
Availability Certain Regions only Why it matters Contact-centre teams can reuse the same MCP tools their agents use elsewhere, rather than building separate integrations.
Vercel Connect your coding agent to models on Vercel's AI Gateway Vercel's setup command configures coding agents such as Claude Code, Codex and Cursor to call models through AI Gateway, including the newly added LongCat 2.5. It needs a free Vercel account and the latest Vercel CLI, and model usage is charged at provider prices.
Step 1
Run Vercel's setup command in your terminal
npx vercel ai-gateway setupCopy Step 2
Select this model name in your agent
meituan/longcat-2.5Copy Image: The Hacker News IBM has disclosed two critical flaws in Langflow, the open-source agent builder, rated 9.8: one bypasses authentication, the other runs arbitrary code without an account. A public Metasploit module already cites the bypass, and IBM named 1.12.4 and 1.12.5 as fixes for 29 further flaws.
Severity 9.8 (Critical), two flaws
Account needed No
Public exploit code Metasploit module cites CVE-2026-93680
Related fixes 1.12.4 and 1.12.5 for 29 other flaws What to do Find any Langflow instances you run, keep them off the public internet and update to the latest fixed release. If one was exposed, rotate the credentials stored in its flows.
Image: Security Affairs Cisco Talos has analysed VOLTA, a self-deleting Windows infostealer that takes passwords, cookies, wallet seeds, OAuth tokens and cloud keys, then sends a profile of the victim to a language model to judge their worth. It tries Claude Opus 4.6 first, falling back to Kimi K2 and Gemini 2.5 Flash.
Platform Windows
Data taken Browser passwords, cookies, payment cards, SSH and cloud keys
Models called Claude Opus 4.6, then Kimi K2, Gemini 2.5 Flash
Analysed by Cisco Talos What to do Avoid keeping SSH and cloud keys in plain files on workstations, and revoke sessions and tokens quickly on any machine suspected of infection.
Palo Alto Networks' Unit 42 describes a campaign, tracked as CL-UNK-1211, against government Linux servers using PHP web shells and a Go backdoor called SysGoRAT. Its scripts carried conversational, step-by-step reasoning notes, which researchers say suggests they came straight from AI output or an automated agent.
Tracked as CL-UNK-1211
Backdoor SysGoRAT, written in Go
Command channel WebSocket to 104.194.158[.]65:443 What to do Check web roots on Linux servers for unexpected PHP files, and search network logs for traffic to the published command-and-control address.
OpenAI says it is more than quadrupling the average midpoint of its reward ranges for medium, high and critical findings. Critical bugs in ChatGPT, Codex or the API Platform now earn up to $30,000, exceptional ones up to $172,000 with new quality and loyalty bonuses, applied retroactively to unpaid reports.
Critical, core product Up to $30,000
Exceptional, with bonuses Up to $172,000
Bonuses Report quality, loyalty, priority campaigns
Pending reports New ranges apply What to do Researchers with pending reports need do nothing to receive the new ranges; anyone planning work should read the updated scope and quality criteria first.
A Northeastern University survey of employers finds AI tools now help screen résumés for 54% of hiring decision-makers. Some are cutting back on beginners: 26% said they were hiring fewer entry-level candidates, against 7% hiring more. About 7 in 10 said more AI training would help.
AI screens résumés 54% of hiring decision-makers
Hiring fewer entry-level 26%
Hiring more entry-level 7%
Want more AI training Around 7 in 10 Why it matters Teams that rely on graduate intakes should check how AI screening and fewer junior roles affect their future pipeline of experienced staff.
10 How companies work with agents Engineering account OpenAI rebuilt its Git backend because agents multiplied the work on its monorepo. OpenAI says Codex let so many people and agents work at once that standard Git hosting became a bottleneck, so it built Pando, a Git-compatible backend, in 10 weeks with Codex. The company says Pando cut median push latency by nearly 80 percent without changing how engineers and agents use Git.
Measure repository load. Track pushes, fetches and merge failures as agent use grows, so bottlenecks show before they stall teams.Keep the interface. Change infrastructure behind the standard Git protocol and client rather than asking people and agents to learn new tools.Plan for parallel work. Expect many concurrent branches from agents and design for branches being updated independently.Use agents for the move. OpenAI used Codex to help teams migrate, a pattern worth testing on internal tooling changes.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 Engineering at Meta · testing Meta's engineers explain how Catching Just-in-Time Tests went from research to full deployment. The post covers each stage of the pipeline and the move from human-in-the-loop review to a human-on-the-loop model, where engineers set the criteria and check outcomes in aggregate.
Why it matters For engineering leads who want AI-generated tests on code changes without adding to reviewers' workload or flooding them with weak alerts.
Simon Willison's llm 0.37 adds GPT-6.1 Sol and Terra, commands for managing OpenAI-compatible endpoints, and attachment support for embeddings. GitHub Google's Gemini app now lets all users, including free accounts, edit the documents, PDFs, spreadsheets and code it creates without leaving the chat. ARY News Microsoft has pushed back default usage-based billing for new Microsoft 365 Copilot Business licences sold through CSP to December 1, 2026. Microsoft Learn Zee Entertainment has appointed Tushar Vohra as its Chief AI and Transformation Officer, a role he took up this month in Mumbai. CXO Digital Pulse Simon Willison points out that the US Census Bureau's business survey now asks firms about "Super Intelligence" instead of AI, following Executive Order 14434. Simon Willison 16 Nov
Deadline · Online GitHub will retire Claude Opus 4.8, Claude Opus 5 and GPT-5.6 Terra across all Copilot experiences and suggests newer replacements for each.
16–20 Nov
Conference · Berlin Five days of training for enterprise software teams, covering developer tooling, legacy modernisation, AI security, quality and testing.
20 Nov
Conference · Zurich A one-day conference on building software with AI agents and on engineering practice for agent-based systems.
25–26 Nov
Conference · London A two-day conference in London, with keynotes and panels from people across the AI industry.
2 Dec
Rules take effect · Online Providers of generative AI systems placed on the EU market before 2 August 2026 must meet the Article 50(2) duty to mark AI-generated content and make it detectable.
