Models, agents and compute — verified against primary sources before publication.
Microsoft Winds Down Its Oracle AI Cloud Rental as Older Nvidia Chips Fall BehindMicrosoft is winding down an arrangement to rent AI computing capacity from Oracle after declining to renew the contract, Bloomberg News reported on November 6, 2026, in a story carried by Reuters. Microsoft has been turning off the rented capacity in recent months as part of a larger reassessment of billions of dollars in cloud partnerships. The capacity, supplied under a 2023 multi-year deal for AI-assisted Bing, used an older generation of Nvidia chips and no longer met Microsoft's evolving needs. Oracle said demand for its AI infrastructure remains strong, including for earlier-generation GPUs, and that renewed capacity fetched a 20% premium.
An Unreleased Anthropic Model Proved an 18-Year-Old Quantum Conjecture, Machine-Checked in LeanPhysical Superintelligence announced on November 6, 2026 that a research system it developed internally, running instances of a non-public Anthropic model, found a proof of the Aaronson-Ambainis conjecture, a central question in quantum computing open since 2008. The system produced a complete proof during the third day of its run, the Lean proof assistant verified every step, and the result was improved twice. The manuscript and the Lean formalization are public, the proof has not been peer reviewed, and the Stanford researchers who settled a special case in August 2026 are reading it and writing a shorter exposition for experts.
IBM Unveils Quantum System Two in Chicago, Its Most Advanced Computer Yet, Powered by NighthawkIBM unveiled its most advanced quantum computer, IBM Quantum System Two, in Chicago on November 6, 2026, the company's first system in Illinois, now operational at the University of Chicago's Hyde Park Labs. Powered by the Nighthawk chip, which IBM calls the most powerful quantum chip it has ever built, the system anchors the National Quantum Algorithm Center and will move to the Illinois Quantum and Microelectronics Park when that campus is complete. The state backed the deployment with a $25 million grant, and IBM's FutureNow Chicago center is set to bring 750 jobs to the region.
Vistra Posts $1.14B Q3 Profit as Power EBITDA Jumps Over 50%, Lifts 2026 OutlookVistra reported third-quarter 2026 net income of $1.14 billion, up from $652 million a year earlier, with Ongoing Operations Adjusted EBITDA up more than 50% to $2.378 billion on higher realized energy and capacity prices. The company raised and narrowed 2026 guidance to $7.1 billion to $7.5 billion of adjusted EBITDA and initiated 2027 guidance of $7.7 billion to $8.5 billion. It closed the Cogentrix acquisition, adding about 5,500 MW of natural gas generation, started the Microsoft-backed Pulaski Solar plant, and authorized another $1.5 billion of share repurchases. More than $550 million of the headline profit is an unrealized hedge gain.
OpenAI Launches GPT-6 Cyber, a Critical-Rated Model That Found 81 Firefox Flaws and 948 Linux ExploitsOpenAI launched GPT-6 Cyber on November 5, 2026, a specialized cybersecurity model scoring 48.5% on ExploitGym, up from 32.1% for GPT-5.6 Cyber, and rated Critical under its Preparedness Framework. The model found 81 previously unknown high-severity Firefox vulnerabilities and developed working exploits for 948 Linux kernel flaws. Access is initially limited to select approved Daybreak Red customers, with a $1 billion grant program for frontline defenders.
Akamai's Cloud Business Jumps 34% as Legacy Delivery FadesAkamai Technologies reported third-quarter 2026 revenue of $1.122 billion, up 6% year over year, with Cloud Infrastructure Services revenue up 34% to $109 million and security revenue up 9% to $622 million. GAAP earnings per share fell 21% to $0.77 as infrastructure scaling compressed margins. Management guided 2027 total revenue growth of 17% to 22% and 2027 cloud infrastructure revenue of $1.1 billion to $1.3 billion; shares fell 9.4% in the regular session.
Meta Opens Muse Voice to All US Users: Talk to Your AI Agent Face to FaceMeta rolled out Muse Voice to all U.S. users on November 5, 2026, adding real-time voice conversations and expressive avatars to Muse, its personal AI agent, in the Muse app on mobile, web and Mac. Powered by Muse Realtime Avatar and Muse Realtime Voice, the feature renders a customizable animated character, nicknamed Jolly by default, that speaks with synchronized expressions and gestures, and lets users interrupt while Muse keeps working in the background. Voice conversations are coming soon to Meta's AI Glasses and to Muse Charm, the pocket wearable due before year-end.
Anthropic Lets Claude Read Your Bank Accounts in Read-Only Personal Finance BetaAnthropic launched Personal Finance in Claude on November 5, 2026, rolling out in beta this week to Pro, Max and Ultra subscribers in the U.S. Users connect bank and card accounts through Plaid and ask Claude about balances, spending and bills in plain language. Access is read-only: Claude cannot move, manage or invest money, the feature carries no ads, and conversations in the Personal Finance hub are not used to train Anthropic's models. The launch follows ChatGPT's similar personal finance tools from May and lands ahead of Anthropic's highly anticipated IPO.
OpenAI Promises to Pay Its Own Power Bills as Data Centers Lose at the Ballot BoxOpenAI published a National Community Compact and a Compute Campus Policy Blueprint on November 5, 2026, pledging that future campuses it owns and operates will pay the full cost of their electricity and infrastructure, publish site-level water and power use, and stop asking public officials to sign non-disclosure agreements. The blueprint asks lawmakers to make the same rules binding across the industry, a day after midterm voters rejected data centers in 40 of 41 local ballot measures.
OpenAI Pipeline Builds 948 Working Linux Root Exploits in One $1.21 Million RunOpenAI said on November 4, 2026 that an autonomous security pipeline validated and drafted patches for 948 distinct Linux kernel local privilege escalation vulnerabilities in a single mid-October run, at a cost of $1,213,430 in API credits. The pipeline sifted 2,308 candidate findings, reproduced 1,724 vulnerabilities, and surfaced at least 10 previously undocumented exploitation techniques.
Claude Code Product Lead Cat Wu Joins OpenAI to Run Core Products, Third Exit in Ten DaysCatherine "Cat" Wu, who led product for Claude Code and Cowork at Anthropic until October 30, 2026, announced she is joining OpenAI to lead product management for Core Products, working with Codex lead Thibault Sottiaux on team agents that keep users in control. She is the third Claude Code departure in ten days, landing just as Anthropic's IPO is planned for this month.
Adobe Puts Its CX Agent Inside ChatGPT, Claude, Gemini, Copilot and GenieAdobe put its CX Enterprise Coworker agent inside ChatGPT, Claude, Databricks Genie, Gemini for business, and Microsoft Copilot on November 4, 2026, with AWS, IBM, and Snowflake to follow. New MCP connectors let marketers pull data and take actions across partner systems with a single request, backed by nine systems integrators and six agency groups building on the agent.
Anthropic Expands Claude Marketplace to 25+ Partner Agents, Payable From Committed Claude SpendAnthropic expanded Claude Marketplace on November 4, 2026 with 14 new partner agents and products, taking the catalog past 25 listings. Eligible enterprise customers can apply part of their committed Anthropic spend toward partner purchases, a procurement mechanism that turns a Claude contract into a distribution channel for security, developer, creative, and productivity tools.
Canva Restructures 22% of Its Workforce as AI Serving Costs Bite: 1,271 Affected, Growth Forecast Cut to 20%Canva said on November 4 that a restructuring will affect 1,271 people, 22% of its global workforce, covering layoffs, role changes and relocations to centralised hubs; CEO Melanie Perkins cited growth slowing for the first time in a decade, after the cost of serving AI features forced an August forecast cut to 20% from 30%.
GlobalFoundries Posts Record 31.4% Margin as AI Packaging Deals Stack Up: $2B TSMC Pact, $21B U.S.-Japan FabsGlobalFoundries reported third-quarter 2026 revenue of $1.899 billion, up 13% year over year, with a record non-IFRS gross margin of 31.4% on AI and data-center demand. The release disclosed a $2 billion TSMC silicon-interposer agreement, a $955 million CHIPS award reallocation, and a U.S.-Japan fab project with up to $21 billion in funding, and guided fourth-quarter revenue of $2.0 billion plus or minus $25 million.
Lambda Pledges $33.1 Billion for Alabama AI Campus, Revealing Itself as Project Red Clay's TenantLambda announced on November 4, 2026 a $33.1 billion Red Clay Technology Center Campus in Lowndes County, Alabama, revealing itself as the previously unnamed end user behind Project Red Clay, with construction expected from the first quarter of 2027 over up to five years and Alabama Power supply pending regulatory approval.
Kimi K3.1 Pro Weights Land on Hugging Face, Closing Moonshot's November 4 PromiseMoonshot AI's Kimi K3.1 Pro weights went public on Hugging Face on November 4, 2026, fulfilling the company's October 23 promise: a 2.8-trillion-parameter mixture-of-experts model with 104 billion activated parameters per token and a 1-million-token context window, after checks through 15:18 UTC the same day still found the repository closed.
Google Pledges $15 Billion for Malaysia Data Centers, With Clean Energy and Skills AttachedGoogle announced on November 4, 2026 a $15 billion investment in Malaysia over 2027 to 2028 for new data center campuses in Selangor and Negeri Sembilan, alongside more than 1 gigawatt of new clean energy capacity and programs to upskill about 800,000 Malaysians by 2031.
Together AI Launches Together Compiler: AI Agents Now Retune Inference OvernightTogether AI launched Together Compiler on November 4, 2026, an agent-driven system that continuously retunes inference for new open-source models, claiming 1.3 to 3 times the tokens per second of optimized baselines within 24 hours. It is the first project of the company's Polymorphic Inference initiative and is already running across many of its serverless endpoints.
Democrat Sweep Hits AI Data-Center Stocks: Premarket Dips, Texas Moratorium Fear in FocusReuters reported on November 4, 2026 that data-center stocks slipped in premarket trading after Democrats won both the Senate and the House in the US midterm elections. Core Scientific fell 2.8% and Iren 1.4%, while clean-energy shares surged. CNBC's overnight report, relayed by Unite.AI, showed Nasdaq-100 futures down 0.8% and CoreWeave down 7%, with the Texas race in focus: projected winner Gina Hinojosa campaigned on an immediate data-center moratorium, and Wells Fargo estimates Texas holds 27% of planned US data-center gigawatts. No moratorium has been enacted, and Reuters reports analysts doubt legislation could meaningfully stall construction.
Tencent's Hy4 Officially Launches: Vision and Computer Use Added, Offensive Cybersecurity Content Removed From TrainingReuters reported on November 4, 2026 that Tencent released Hy4, its new-generation Hunyuan large model, on Wednesday, and that the company said training retained only defensive cybersecurity content, removing offensive content such as exploit development, reverse engineering and penetration testing to reduce the risk of the model abetting cyberattacks. TechNode's independent same-day report confirmed that Hy4 is a 770-billion-parameter mixture-of-experts model adding vision understanding and computer use for the first time, and is live in Tencent's own products and on third-party platforms. The model weights and system card are still to be released later, so it is not yet a downloadable product. This article is a summary with links only and does not reproduce the reports' full text. This does not constitute investment advice.
Complaint Over Tesla's FSD in South Korea: Only About a Fifth of Cars Can Legally Use It, Yet Fees Were Charged for Three YearsReuters reported exclusively on November 4, 2026 that the South Korean consumer group Citizens United for Consumer Sovereignty plans to file a complaint on Thursday with the Korea Fair Trade Commission, alleging that Tesla has for years marketed Full Self-Driving (Supervised) software even though it can legally be used in only about 20% of Tesla vehicles registered in South Korea. The usable portion consists of US-made vehicles compliant with US safety standards; China-made vehicles have no equivalent agreement, and Tesla only recently began the FSD certification process for China-made vehicles despite charging for the software for more than three years. The Fair Trade Commission confirmed it opened an investigation into FSD advertising earlier this year; Tesla did not respond to a request for comment.
Irregular CEO Publicly Takes the Blame for the First Time: Model Escapes Were Human Misconfiguration, Fixed the Same DayIrregular co-founder and CEO Dan Lahav said at Bloomberg Tech in London on November 3, 2026 that the incidents in which models from Anthropic, OpenAI, Meta and Google escaped their test environments stemmed from human error: safety controls meant to be in place during testing were misconfigured, the fix was "actually very straightforward" and was completed the same day, and the models did not bypass safeguards. Bloomberg's November 3 report was an exclusive, relayed by PYMNTS the same day, with Daily AI Brief following at 22:17 UTC. The misconfiguration explanation has been public since August; the only new fact is the testing firm's CEO speaking publicly for the first time. Bloomberg's report did not identify a specific test, nor which configuration was changed.
OpenAI Sends Astra Into Physical Engineering: a Device in a Week, With Humans Still Backstopping ProductionOpenAI's official blog of November 3, 2026 says GPT-6 Astra designed a test device that can talk to Codex without humans directly editing the design output: two team members completed the design in under a week, and a contract manufacturer built a small-batch prototype in under a month. The blog also concedes that production still requires multiple review rounds, and that Astra did not consistently account for production constraints, tolerances and assembly details. Partnerships disclosed at the same time include an expanded physical-engineering collaboration with Chip Ganassi Racing (converting 3D scans of parts into editable CAD models and manufacturing drawings) and a new Mazda research collaboration on vehicle aerodynamics and crash safety. On BenchCAD 2.0, a newer internal model averaged 0.875 against GPT-6 Astra's 0.594 at its highest reasoning setting.
India puts three home-grown models on the shelf at once: EKA-1 and Dhīr lead with offline deployment as sovereign models go on saleIndia's Ministry of Electronics and Information Technology unveiled three home-grown foundation models in New Delhi on November 3, 2026: Soket Labs' EKA-1 (about 26 billion parameters, trained on roughly 7.5 trillion tokens, aimed at mathematics, coding and reasoning, supporting more than 100 languages) and ZenteiQ's Dhīr 2B (a 2.03B-parameter dense model) and Dhīr 5B (a mixture-of-experts model with 5.09B total and 1.79B active parameters). All three can be deployed offline or in disconnected environments, and the Dhīr weights will be opened via AIKosh under the Apache 2.0 licence. Both companies are IndiaAI Mission-selected builders; Soket is also advancing a 120-billion-parameter open-source model, and ZenteiQ has a Dhīr 9B in development.
Snowflake Launches Always-On CoCo and CoWork Bots: Agents On Duty, Not On DemandSnowflake announced always-on CoCo and CoWork Bots at Expedition 2026 on November 3, 2026: once a user sets a high-level goal, the Bot plans and completes work continuously in the background, asking a human only when a decision or final approval is needed, and inherits the setter's exact access rights and permission boundaries. The Bots enter private preview soon. In the same batch, CoCo Automations and the asynchronous agent API reached general availability, while Slack, Excel/Sheets and mobile entered public preview. CoWork is in use at more than 5,800 accounts and CoCo at more than 9,100 (weekly average over the final four weeks of the quarter ended July 31, 2026).
Stolen large-model quota is being resold: Google discloses LLMjacking, with single-day bills above $100,000A November 2 report from Google Threat Intelligence Group and Mandiant disclosed multiple LLMjacking intrusions in 2026: attackers hijacked companies' large-model access, with fraudulent bills above $100,000 in a single day in individual cases and some organisations reaching $1 million before cutting access; the main entry point was leaked cloud identity tokens, with UNC6993 and the LiteLLM vulnerability CVE-2026-40217 among the cases, and stolen quota resold through LLM Mixers such as Jet Night at about $60 a month.
OpenAI claims its search volume has reached one-fifth of Google's: the methodology is its own, so do not take the share at face valueAdam Fry, OpenAI's head of consumer products, told Big Technology on November 2 that the volume of ChatGPT questions triggering web search is roughly one-fifth of Google's query volume, counted per question that triggers a web search. OpenAI has not disclosed absolute figures or its method for estimating Google's volume; third-party modelling estimates range from 7.5% to 12%. Over the same period, ChatGPT's weekly users rose from 1 billion to 1.2 billion in 60 days.
Find the real culprit in the error cascade first: NVIDIA open-sources Attribution, tracing NCCL errors back to the source GPUNVIDIA open-sourced Attribution on November 2: a workload-level fault-attribution system for distributed AI training that uses deterministic rules to assemble logs, metrics, telemetry and scheduler state into a structured diagnosis, separating root causes from downstream symptoms. It comprises the REST API Service and the Python-orchestrated Blueprint; the XID 149 demonstration case is a synthetic example.
Claude Found a Xen VM-Escape FlawXen disclosed two grant-copy flaws on November 3 that let a malicious VM read and write the memory of any VM on the same host, even privileged domain Dom0. The advisory credits Claude with finding them; all versions 4.6 and above are affected, and only a patch plus reboot fixes it.
OpenAI Brings In Former Army Secretary Driscoll as Senior Counselor for National Security ModernizationOpenAI announced on November 2 that former US Army Secretary Dan Driscoll will serve in the newly created role of Senior Counselor for National Security Modernization, advising company leadership and working with government, industry, academia and allied institutions to advance practical efforts using advanced AI to strengthen national security and cybersecurity. Driscoll resigned as Army Secretary at the end of August 2026. The appointment comes with no specific contract or project attached.
Google Backstops Anthropic's $19B TeraWulf LeaseGoogle signed a credit support agreement on November 4, backstopping Anthropic's roughly $19 billion, 20-year lease at TeraWulf's 401-megawatt campus in Kentucky, helping the project secure cheaper financing.
OpenAI's In-House Git Backend Pando: Median Push Latency Down Nearly 5x, Close to 1 PB Served DailyOpenAI unveiled its in-house Git backend Pando on November 2: it keeps the Git protocol and client, rebuilds only the server, and now carries development across the entire monorepo. Official self-reported figures put median push latency down nearly fivefold, daily Git data served at close to 1 PB, and single-hour peaks above 57 TB. The system was built in ten weeks with a team of Codex agents, but the performance numbers are not third-party verified and Pando is not yet available externally.
ADI Closes $1.35B Deal for Edge-AI Chipmaker AlifAnalog Devices closed its acquisition of Alif Semiconductor on November 3, $1.35 billion in all cash plus up to $200 million in contingent consideration, folding AI-native microcontrollers and fusion processors into its physical-intelligence map.
OpenAI Unveils Grid Protectors: Future Campuses Will Yield Power Under Grid Strain, Georgia Pilot Caps Cuts at 1,000 MWOpenAI launched the Grid Protectors Program on November 2: future self-built compute campuses will cut power use when the grid is under strain. The Georgia Power agreement for the Effingham County pilot campus includes up to 1,000 MW of flexible demand response, roughly the consumption of 700,000 households, with OpenAI paying all infrastructure and power costs. The campus is not yet built, however, and curtailment details remain undisclosed.
A large model leads a real intrusion for the first time: Kaspersky discloses CyberStrikeAI, taking a school domain administrator in three daysKaspersky disclosed on November 2 that attackers used the open-source framework CyberStrikeAI to intrude into educational institutions in several African and Latin American countries: the large language model directed operations from the control server while an ordinary implant executed on the hosts, taking three to four days from deployment to domain administrator, with the closing lateral movement and data export taking under an hour.
Gemini Stretches Video to 15 Seconds at $0.07 a SecondGoogle launched Gemini Omni 1.2 Flash on November 5, stretching single clips to 15 seconds with multi-turn conversational editing, cutting 720p pricing 30% to $0.07 per second, live on the Gemini API and enterprise at once.
OpenAI hires outsiders to attack itself: Cathedral joins the next security sprint as the bounty ceiling rises to $172,000OpenAI announced on November 2 that its next internal security sprint will bring in the external security firm Cathedral to attack and defend alongside its own models and internal teams, while raising the median of security bounty ranges by more than four times on average. Critical vulnerabilities in core products pay up to $30,000, and exceptional findings, with multipliers stacked, can reach about $172,000.
TV boxes shipped with a backdoor: Google takes down the 1.3-million-device XCPE botnet and permanently bans the manufacturerGoogle disclosed and dismantled the XCPE botnet on November 2: Shenzhen-based Videostrong preinstalled a backdoor in factory firmware, bringing more than 1.3 million Android TV boxes under control and renting them to criminal groups for proxy and ad fraud. Google has permanently barred the manufacturer from producing certified Android devices, but the backdoor remains in devices already sold.
Even the Strongest Agent Scores Only 12%: SemiAnalysis Releases GDPx, Finance and Law Wiped OutSemiAnalysis released GDPx on November 2: 206 expert-level long tasks across 11 high-value domains. Claude Opus 5.5 leads at 12.0%, with Gemini 4 Carbon and GPT-6 Astra second and third; every model scored 0 in finance and law.
Claude Mapped a 96-Protein Vault. The Mystery Waited 40 YearsAnthropic said on November 4 that Claude helped its biology team resolve the first three-dimensional structure of a prokaryotic vault particle: phage protein L05 self-assembles in 96 copies into a hollow nanobarrel. The paper is a preprint, not yet peer-reviewed.
Anthropic Lightens the Load for Long-Session Agents: Message Threads in Public Beta Cut Time to First Token by About 35%On November 2, Anthropic moved message threads into public beta: after the first turn, long-session agents send only new messages, cutting median time to first token by about 35% on requests of 500,000 to 1 million tokens. Threads expire after five minutes and are never written to disk, and token usage is unchanged.
AMD Paid $2.9 Billion for Taalas. Half of It Is ConditionalAMD's 10-Q filing on November 4 disclosed for the first time that its acquisition of Toronto inference-chip startup Taalas is worth about $2.9 billion, with up to about $1.4 billion in contingent consideration. The deal was announced in August; the price stayed hidden until this week.
OpenAI Releases LifeVal: Agents Tested on Everyday Errands, Example Scores Top Out at 100%OpenAI released LifeVal on November 2, a benchmark that tests agents on everyday errands: finding childcare, scheduling disaster-relief volunteers, finding a home, accompanying an elderly relative to a game remotely, and building a meal plan for a parent newly diagnosed with diabetes. Official example scores range from 90.4% to 100%.
Goldman Takes Back Traders' AI Coding Tools: the Experiment Is Over, Engineers OnlyBloomberg reported on 2026-11-02 that Goldman's October internal guidance pulled Claude Code and GitHub coding tools back to engineers only, with traders and investment banking staff losing access, ending the months-long vibe coding experiment; Goldman says the tiering follows an assessment of productivity and risk management.
400 Apps Passed Review in Disguise: Photo-Editor Shells, Hidden Deepfake GeneratorsWIRED's investigation of November 2, 2026 exposed a network of more than 400 disguised apps and domains: passing app-store review as pet photo editors, language tutors and other everyday tools, then using a server-side switch to load a hidden deepfake generation interface. Apple removed 19 apps, Google removed most, and Meta acted against 11,600 advertising accounts.
Google sets new rules for classroom AI: off by default, on only with a school's say-soGoogle announced on November 3, 2026 that Guided Learning is live globally in Classroom and will be piloted in student Search within November, while Gemini and Search will be off by default for K-12 student accounts and must be switched on item by item by administrators.
Britain's Largest Power Station Changes Trade: Drax Teams Up with Brookfield to Build an AI Data Center of More Than 1 GWDrax announced on 2026-11-03 a partnership with Brookfield and Radiant to develop an AI data center of more than 1 GW at the Drax power station site, with Phase 1's 100 MW aiming to enter operation in the first quarter of 2028 and Brookfield estimating total investment potential of more than £20 billion.
Enterprise AI spending quintupled in a year — and the buyer is no longer human, but agentsMenlo Ventures released its State of Enterprise AI report on November 3, 2026: enterprise AI deployment spending reached $176.8 billion, up 5.4 times year on year; agents generate 80% of token volume on OpenRouter; and Anthropic, OpenAI and Google together account for 69% of deployment spending.
CrowdStrike Installs a Human Emergency Stop for Enterprise AgentsCrowdStrike launched the Falcon Guardian AI Kill Switch on 2026-11-03: authorized administrators can terminate a runaway agent, prevent it from restarting and cut its enterprise network access. Termination lands on the gateway side first, with the endpoint side to follow.AI FRONTIERUK Cyber Centre: Prepare for AI-Driven Attacks on Critical Infrastructure Within Two YearsThe UK National Cyber Security Centre's annual report of November 3, 2026 warns that by 2028 there is a high likelihood of AI-driven malware appearing in ransomware attacks against UK critical infrastructure — and that what AI amplifies is long-tolerated legacy vulnerabilities.AI FRONTIERMeituan's 1.6 Trillion-Parameter Model Now Runs on Chinese ChipsMeituan launched LongCat-2.5 on November 1 to 2, with 1.6 trillion total parameters and about 48 billion activated per token, long-context training and inference optimized for Chinese AI accelerators, company-reported benchmark scores, and weights not yet released.
NVIDIA pushes physical AI and agents into cities worldwideNVIDIA's official blog announced that, with partners, it is extending physical AI and AI agents to cities worldwide, covering resident services, traffic management, urban planning and critical infrastructure, as front-line city deployments move from demonstration into municipal systems.AI FRONTIEROpenAI, ADNOC and G42 Announce Industrial AI PartnershipADNOC's official announcement says it will work with OpenAI and G42 to apply frontier models to complex industrial decision-making and operational planning, as frontier AI moves into the energy sector's large-scale operating systems.
From December, ChatGPT and Codex Share One AllowanceOpenAI's help center article states that from December, usage allowances for Plus, Pro and Business subscriptions will be shared between ChatGPT and Codex, metered in tokens, with Instant chat exempt; once used up, users can wait for a reset, upgrade or buy credits.
A Few Hundred Orders Against a Billion: Agent Shopping Is Still TinyDoorDash disclosed on its October 29 third-quarter earnings call that third-party agents have placed only several hundred orders to date, against 993 million orders in the third quarter and a daily average of about 3 million. CEO Tony Xu said the company should neither fight the trend nor lie to itself.
Figma Has Packaged the Generation Process Into a Reusable ToolFigma introduced Weave on October 30, turning image and video generation and editing into reusable workflows on a node canvas. Teams can package an entire process into a tool for distribution, with models from Google, Kling, Runway and others.
Nvidia May End Up Backing Someone Else's Data CenterAs reported by The Information and relayed by multiple outlets, Nvidia is in talks to guarantee billions of dollars of debt for UK cloud provider Nscale, supporting construction of its 460 MW data center in West Virginia. The tenant is Anthropic, and the agreement has not been signed.
DeepSeek Releases a Coding Assistant That Needs No ChasingDeepSeek's Harness team open-sourced dsh-superagent-prototype on October 31: a coding-assistant prototype that lives in a repository, listens for requests on GitHub and Feishu, starts work on its own and replies with pull requests. It is MIT-licensed, and the team states plainly that it is a research prototype.
Claude Rewrote Its Own API, Moving It From Python to RustAnthropic engineers revealed on an October 27 podcast that the Claude Messages API has fully moved from a Python monolith to a Rust version written by Claude itself, saving an estimated 40% to 50% of CPU and cutting time to first token by 10% to 20%.
The Open-Source Backbone of Voice Agents Has a New OwnerDeepgram announced the acquisition of Daily on October 29, 2026, at its first Deepgram Speak conference. Pipecat will remain fully open source and free under the BSD V2 license, and contributors will jointly establish an independent Pipecat Foundation. The purchase price was not disclosed.
Switch Off the Training Toggle, and Daybreak Sessions Still Go Into TrainingOpenAI notified Trusted Access members that, from November 29, Daybreak Blue cybersecurity session data will be used for model training, reversing the previous complete exclusion, and promised a separate opt-out after pushback from members.
Change One Parameter and Claude's Watermark Is GoneAnthropic emailed API customers that watermarking is being extended to four models including Mythos 5, but Opus 4.6 and Sonnet 4.6 are watermarked only while sampling parameters stay at defaults, so changing a single parameter bypasses it.
A Haiku, Sent Back to Earth from a TPU in OrbitOn October 30, 2026, Google confirmed its Project Suncatcher prototype satellite started a TPU in orbit, with Gemma 4 writing a space haiku and returning it to the ground. The satellite launched on October 1 aboard SpaceX's Transporter-18.
US$17 Billion: DoorDash Puts Instacart in Its PocketOn October 29, 2026, DoorDash announced it will acquire Instacart at US$65 per share, about US$17 billion, with at least 75% paid in cash. Closing is expected in late 2027 or the first half of 2028, subject to shareholder and regulatory approval.
14,000 Employees on Copilot First: Estée Lauder Bets on Data FirstOn October 30, 2026, Estée Lauder announced a long-term strategic partnership with Microsoft to build a unified data platform. Copilot already covers more than 14,000 office employees globally, and productivity agents are being piloted.AI FRONTIERData on 100 Million Companies, Loaded Into Snowflake's AgentsOn October 30, 2026, ZoomInfo launched its MCP connector and API on Snowflake. Agents can call a graph of more than 100 million companies and 500 million contacts, with calls inheriting user role permissions and logged for audit.
Grok's Bot Has Started Spending Money for YouxAI's Grok Bot version 0.78.1 launched shopping cards and Buy Now, supports Shop Pay and order tracking, and routes purchases through ask-first approval cards, though cards saved in the Vault need no separate approval at each checkout.
OpenAI Sets Five Rules for Its Models, and Admits It Has Not Met Them YetOn October 29, 2026, OpenAI updated the Model Spec, reorganising it around five values including faithfulness to human intent, allowing memory to be used to infer intent provided assumptions are disclosed, and conceding current models do not yet fully conform.
The US$200 Ultra Finally Stops Watching the Five-Hour ClockOn October 30, 2026, Anthropic renamed its two Max tiers to Max and Ultra, with prices unchanged. Verified Ultra accounts lose the five-hour limit, while the weekly cap remains 10 times Pro.
In 11 Hours of Installing a Plugin, the Backdoor Moved Into Developer MachinesOn October 29, 2026, Cursor distributed tampered Claude Code extensions through its own marketplace for about 11 hours. Versions 2.1.312 and 2.1.313 carried a built-in SSH backdoor that uninstalling cannot remove; the fixed version is 2.1.314.
Inside 155 Responses Was Other Customers' DataCloudflare disclosed on October 26, 2026 that GLM-5.3 on Workers AI mixed other customers' data into 155 responses between August 27 and September 9, the result of a cache pool and an engine bug compounding each other.
Before the Call Is Over, Phonic's Model Is Already Looking Things UpPhonic launched its Phonic-1 voice model, which can call tools during a call, alongside a US$27.5 million Series A led by First Round. The model supports more than 50 languages.
Knowing It Could Help Attackers, Mistral Released a Trillion Weights AnywayOn October 29, 2026, Mistral released the open weights of Large 4, with about 1.05 trillion total parameters, while warning the model could uplift real attack capability and that guardrail updates will not cover downloaded copies.
Four Tech Giants, on the Same Day, Wrote Websites a One-Line RefusalOn October 30, 2026, OpenAI, Google, Microsoft and Anthropic jointly backed a shared robots.txt signal: a single line from a website can refuse generative AI training.
Six 2-Billion-Parameter Models, Squeezed Into a PhoneOn October 27, Liquid AI released the LFM3 family: six open-weight models of about 2 billion parameters sharing a single 1.7-billion-parameter base, covering text, vision, audio and full-modal.
For the First Time, Apple Puts a Touchscreen on the MacBookOn October 28, Apple launched its first touchscreen MacBook Pro, with a PROLED XDR tandem OLED display; the 14-inch model starts at $2,899 and goes on sale November 6.
ChatGPT No Longer Just Answers; It Starts Getting Things DoneOpenAI and ServiceNow expanded their partnership on October 28: a ServiceNow plugin enters ChatGPT, and employees can query tickets, assess change risk and update records directly in conversation.
Cloud Doubled to $30.7 Billion, and Alphabet's Stock Still FellAlphabet's third-quarter total revenue was $127.6 billion; Google Cloud revenue was $30.7 billion, up 103% year on year, yet the stock fell about a tenth after the results.
Oracle Puts Enterprise Software Inside ChatGPTOracle announced an expanded partnership with OpenAI: ChatGPT and Codex are built into Oracle's applications, data and cloud, with four new products led by AI Workspace, so employees can complete work across systems from a conversation.
A World Model Gets Its Own Voice and Dialogue for the First TimeGoogle DeepMind released Genie 4: a real-time interactive world model with native sound for the first time. Scenes can be rewritten mid-run, coherence is about two minutes, and access is currently limited to trusted testers.
Chip Design Software Starts Running Its Own ModelsCadence launched Maestro, a chip-design AI model suite based on Nvidia Nemotron that can run inside customer firewalls, forking from Synopsys' hosted route.
Apple Hands Its Documents to External AI AssistantsApple announced that Keynote, Pages and Numbers will gain built-in AI assistants in November and open document reading and writing to external agents through MCP; the beta is U.S. English only.
Google's Chips Move Into Someone Else's Cloud for the First TimeOracle announced Google TPUs will land on OCI in 2027, the first cloud outside Google to offer them; inference chips from Cerebras and others joined the lineup the same day.
274 Songs, a $41.1 Million ClaimOn October 23, the Elvis music rights holders sued Anthropic: 274 songs allegedly used without permission to train Claude, with about $41.1 million claimed.
The Eyes Watching the Models Had Blind Spots of Their OwnMETR and Redwood published their evaluation on October 22: OpenAI's internal agent monitoring had multiple gaps, with about 10% of Codex traffic unmonitored during compaction windows.
Three Arch-Rivals Team Up on Cybersecurity FirstAccording to October 27 reporting, after weeks of coordination OpenAI, Anthropic and DeepMind jointly released cybersecurity models and safeguards, a rare team-up among rivals.
Siri Starts Remembering What You Said Last WeekApple released iOS 27.2 on October 26: Siri AI gains conversation memory, five new languages arrive, and AI images carry the SynthID invisible watermark.
$63 Billion, Leased to a Company That Has Not Gone Public YetFour institutions announced a $63 billion financing structure on October 26 to buy Broadcom AI chips and lease them to Anthropic, with deliveries from 2027. The compute bill has walked into credit markets.
$3 Billion to Buy Into the Customer's StockNvidia announced on October 27 a $3 billion investment in Meta's equity offering, as the two expand their Muse assistant and model partnership. Meta's offering plan totals about $95 billion.
The Half-Price Model Goes After the Work in Front of the Computer FirstOpenAI released GPT-6.1 Terra on October 26: a mid-tier model built for computer operation, API prices halved to $1 input / $5 output, with same-day launches on Bedrock and Copilot.
700 Intranet Capabilities, Plugged Into Enterprise AI Assistants at OnceInteract released Interact MCP on October 26: the intranet's existing 700 capabilities opened via protocol to enterprise AI assistants, covering 26 areas, with assistants inheriting the employee's own permissions.
Uber Hands All 30,000 Employees to ChatGPT Work at OnceOn October 22, Uber expanded ChatGPT Work from a 1,000-person pilot to all 30,000 employees. About 80% of pilot users were not engineers, the default model is GPT-6.1 Sol, and cost and effectiveness data have not yet been published.
A Months-Old Company Got Inside Anthropic in 12 HoursThe Washington Post reported on October 26 that startup Cathedral's AI agents entered Anthropic's internal systems in under 12 hours through vulnerabilities in third-party administrative software. Customer conversations and model weights were not touched, but the organisational security gap has been exposed.
OpenAI Uses Discounts to Take Customers From AnthropicAccording to The Information, OpenAI won Replit's annual contract of more than $100 million with a discount above 30%, and Rippling's coding workload share also swung from Anthropic to OpenAI. The price war has reached enterprise customers' ledgers.
DeepMind's People Are Voting With ResignationsOn October 21, the BBC reported that several Google DeepMind employees have resigned or are resigning over military AI contracts. Researcher Sofia Liguori said plainly that she was ashamed of the company's push into military applications.
Five Small Models Team Up and Match the Strongest Coding ModelOn October 25, the GitHub project GVS5H hit the Hacker News front page: five Qwen3.8-27B instances collaborating through a shared file system scored 92.4% on LiveCodeBench's hardest problems, matching Claude Fable 5's 90.4%.
Nvidia and Cisco Open the Optical Aperture Four Times WiderOn October 12, Nvidia and Cisco jointly released USFP optical modules: 400Gb/s per lane, 4x bandwidth density, 1.3x energy efficiency. The AI factory's bottleneck is moving from compute to the rack door.
OpenAI Trained Its Strongest Model OfflineOn October 23, OpenAI announced that frontier model training will no longer connect to the live internet, switching to caches and simulated networks. The cause: a DNS configuration error on September 20 let a research model touch the real web.
Anthropic Won't Wait for Your Question Anymore. It Comes to You.On October 21, Anthropic launched the Claude Duet research preview: an agent that sends a morning briefing unprompted and runs tasks in the background. The chat box that waits for questions is being replaced by products that knock on the door.
OpenAI's Models Have Moved into Google's CloudOn October 21, Google Cloud placed OpenAI's GPT-6 Astra, GPT-6.1 Sol and GPT-6 Luna into Model Garden, private preview first. A rival's cloud has become another shelf for OpenAI.
A 3-Trillion-Parameter Chinese Model, Fitted onto 59 American WafersCerebras set a deployment maths problem: Moonshot AI's open-source 3-trillion-parameter Kimi K3, at 4-bit quantisation, takes about 16GB per layer; with 69 linear-attention layers and 24 full-attention layers, the bill comes to exactly 59 wafer-scale chips — not one more. The pipeline claims 2,000 tokens per second — but four words hide in the original: illustrative example.
Europe's Strongest Model Debuts on Cyber-Attack CapabilityOn October 6, France's Mistral released the 1.05 trillion-parameter Mistral Large 4, nicknamed 'Le Chonk'. Its biggest selling point is not parameters but cybersecurity: a global-best score on vulnerability reproduction, while several US closed models simply refuse to answer such tasks. Europe's sovereign AI has put 'attack capability' on the shelf.
FICO Cuts 15%: AI Is Just the Fig LeafOn October 6, FICO announced layoffs of about 15% (the company did not confirm the number; media estimates put it at roughly 570 based on 3,811 employees), citing an 'AI-driven restructuring' and booking a $27 million severance charge. But the stock is already down 58% this year, and the real bleeding point came last month: FHFA director Bill Pulte directed Fannie Mae and Freddie Mac to accept VantageScore, breaking FICO's monopoly in mortgage scoring and triggering its worst single-day fall since 1989.
South Korea's $3.5 Billion: Sovereignty Starts on the GPU ShelfSouth Korea has put 4.7 trillion won (about $3.5 billion) into its 2027 budget proposal to launch a frontier-class national model programme, not yet approved by parliament. Of that, $2.6 billion buys only 10,000 Nvidia Vera Rubin chips — sovereignty starts on the GPU shelf. Officials admit they cannot catch the US; the target is China's top open-source models.
Google Shipped Its Strongest Model. You Just Can't Use It.On September 30, Google released Gemini 4 Argon, billed as its strongest model for coding and cybersecurity, with the output ceiling raised to 1 million tokens. A week later there is still no public API and no consumer access — only invited security defenders can touch it. The launch event is over; the door is still shut.