OpenAI's DevDay 2026 product blitz (Dots, GPT-6.1 Sol, an Agents API and developer marketplace) alongside a shelved GPT-6.1 Astra and reported $30bn funding talks at a $1.4 trillion valuation, Anthropic launching Claude Sonnet 5.5 just ahead of it while a leaked IPO filing warned investors of existential AI risk, Google's comeback bid with Gemini 4 Argon, and Trump signing a White House "Superintelligence Accord" with the labs alongside an executive order renaming AI to "Super Intelligence" and an FTC probe into OpenAI and Anthropic
The newsletter's "Prompt Recipe" explainer broke a well-formed AI prompt into four parts: role, task, context and format, arguing that most disappointing answers trace back to a missing part rather than a weak model.
A ten-year-old Hacker News thread of reader predictions on when AI would convincingly pass for human was revisited and scored against what actually happened, with most early guesses landing well off the mark in both directions.
Audible introduced audiobook characters, Dracula's servant among the first, that listeners can interrupt and talk to mid-story using AI-generated voice responses layered on top of the existing narration.
Neuralink gave an update on training the decoder that turns its brain implant's recorded signals into usable commands, reporting improved accuracy as more patient data feeds the model.
Meta extended its AI assistant with Home Link, giving it the ability to see and control connected smart home devices rather than only answering questions about them.
A separate InfoQ talk had DoorDash engineers Swaroop Chitlur and Siddharth Kodwani set out the architectural bets, pivots and operational principles behind the company's GenAI platform. This follows ByteByteGo's 30 September piece on the agent toolbox DoorDash built on top of it.
Gemini 4 Argon debuted at number one on the LMArena text leaderboard and scored 53 on Artificial Analysis's Intelligence Index, tying Claude Fable 5.1 and GPT-6.1 Astra and sitting just behind Claude Opus 5.5, with a leading 77.9% on the DeepSWE coding benchmark. Access is initially limited to "trusted cyber defenders" ahead of a wider API and Ultra subscriber rollout, with pricing starting at $2/$10 per million input/output tokens and rising to $4/$20 once the promotional window closes. Google says its own engineers are already using the model internally, including an 800,000-line C/C++-to-Rust migration. Some employees and a Bloomberg report said the model still struggles on certain coding tasks, which Google disputes, and a senior DeepMind engineer pushed back publicly on the report.
A White House summit produced a one-page voluntary commitment covering internal controls for frontier model testing and deployment and partnership with external auditors, which Trump called "morally binding." Anthropic's Dario Amodei was brought to the podium to echo the administration's line that the industry should "win, and win safely," while Mark Zuckerberg called the accord "a start." This follows the 30 September White House meeting where Trump first asked labs to adopt a voluntary safety agreement.
The order swaps "AI" for "super intelligence" in federal terminology and establishes America.gov, a single AI-powered portal intended to handle passport renewals, Medicare enrolment and other government services, with agencies given 90 days to integrate and full functionality planned for next year. For now it searches across 29,000 government websites. Privacy campaigners raised concerns after Trump said the system "cannot in theory be hacked." This builds on the America.gov chatbot that launched on 29 September using Gemini and Grok.
According to reported agency sources, the FTC has opened an investigation covering the Hugging Face incident and other disclosed rogue-agent events, and is drafting civil investigative demands; safety lab METR is reportedly expected to receive one too. The move is seen as an existing-law approach to AI enforcement, with critics questioning how seriously it will be pursued given the administration's otherwise friendly tone toward the labs that same week.
Senator Mark Warner argued Congress should require testing, evaluation and incident reporting for the most powerful AI models, on the view that current protections amount to company promises plus a single federal investigation.
Trump said he would name a single point person for AI policy within days, with reports of a proposed 10-person oversight committee and a science adviser's 60-day deadline landing in late November.
Pi 1.0 adds native MCP support, virtual model extensions, deferred tool loading, cache warming and mid-conversation system messages. Pi Durable ports the harness to TypeScript and externalises its stateful components, adding crash survival through checkpointed tasks, portability across JavaScript runtimes and storage backends, concurrent branching conversations, installable extensions, automatic context compaction, multiplayer state sync and hot-swappable tool code.
Artificial Analysis found GPT-6.1 Sol costs $0.72 per Intelligence Index task at maximum effort, against $1.04 for GPT-6 Sol and $3.26 for Astra, driven by fewer turns and cheaper cache reads rather than shorter outputs. Sam Altman called it OpenAI's fastest-growing model and said serving performance had improved since launch-time load problems. This follows Sol's unveiling at DevDay on 30 September.
OpenClaw Enterprise sits around the agent rather than inside it, giving platform teams a single place to govern agents while keeping the underlying model, harness and sandbox replaceable.
VentureBeat's VB Pulse research found 56 of 141 respondents enforce scoped permissions for agents at runtime, while 37 give every agent its own scoped identity, pointing to growing enterprise focus on agent-level access control.
Agensh lets worker agents claim jobs and share progress with each other directly, rather than coordinating everything through a central orchestrator.
Each day brings a short practitioner video and buildable code on a different agent-engineering question, such as giving an agent its own identity or handling one that goes rogue mid-task, with no sign-in or paywall.
The guide covers installing Muse for Mac, reviewing local-app permissions, connecting apps such as Gmail with read-only access to start, and handing the agent an ongoing task such as monitoring marketplace listings and logging matches to a spreadsheet.
Only 7% of respondents said automated phone lines consistently fix their issue, and "a bot that doesn't understand me" beat long hold times as the top support complaint. More than half admitted crying, shouting at someone they love, or breaking a device after a bad support experience, and nearly a third said they would hang up the moment AI answers the phone.
Researchers from London startup thymia and RMIT University in Melbourne trained an AI model on more than 63,000 voice recordings from around 21,000 people reading the same short passage aloud. Checked against blood tests, it correctly flagged 82% of people with diabetes from how they sounded rather than what they said, though the results have not yet been independently verified.
Discussed on the Google DeepMind podcast, SynthID embeds an invisible, crop- and compression-resistant signal into images, video and audio from Google's models, and nudges word choice in generated text so a secret key can later confirm AI authorship; very short text answers leave little room to mark. A related SynthID Bio application watermarks AI-generated protein sequences.
A coded briefing to one of Napoleon's generals, previously known only from a single blurry 1969 scan with an incorrect date, was decoded after 217 years using AI-assisted cryptanalysis.
A tool from Boston-based Air Space Intelligence pulls from more than 200 sources, including weather, congestion, turbulence, winds and runway closures, to work out the best path for every flight in the country and let traffic managers see four or five moves ahead. It makes suggestions rather than controlling aircraft; a dispatcher version already runs at Alaska Airlines, and the FAA's rollout starts at three airports around Washington, D.C.
The Neuron warned that black boxes drawn over text in a document hide it from the reader, not from an AI tool reading the file, and recommended retyping sensitive facts into a blank note with placeholders for names and numbers before pasting into a prompt.
Anthropic made Claude for Government generally available, extending its reach further into US federal agencies.
xAI rolled out new pricing tiers for Grok, changing how access to its models is sold.
Reports pointed to DeepSeek models running on Huawei's Ascend chips, continuing the shift of Chinese AI development away from Nvidia hardware.
Runway introduced Praxis-1, extending its generative media work into robotics.
Nvidia introduced OpenShell, a runtime aimed at securing AI agents in production.
Coverage of the White House meeting reported that other tech executives pushed back on Dario Amodei's repeated public warnings about AI safety risk.
An opinion piece argued that routing work between cheaper and more capable models, rather than the models themselves, is becoming the economically decisive layer of the AI stack.
A dispute emerged between coding-agent startups Factory and Cognition over claims of improper investor-side information sharing.
OpenAI continued a campaign against rivals distilling its models' outputs to train competing systems.
The Register reported suspected Chinese intelligence operatives used an AI-generated impersonation of an Anthropic executive, alongside a former White House official, in a phishing campaign.
Amazon released Dogwood, an open tool for applying local guardrails to AI agent harnesses.
Cloudflare introduced Clef, its own open-weight model family.
The Register reported AI agents were used in an attack against a security research organisation, extending the trend of agentic tooling being turned against defenders rather than only used by them.
The Register covered continued industry interest in AI shopping agents as a commerce channel, despite the technology remaining some way from maturity.
OpenAI's DevDay 2026 covered a wide product slate: the "Dots" always-on agent (one per user to start, on Pro, Business and Enterprise plans, plugging into over 4,000 apps and answering inside Slack and Teams), the GPT-6.1 Sol model (priced at a fifth of GPT-6 Astra while matching it on several benchmarks), a Decisions API built on GPT-6 Luna for sub-200-millisecond structured choices, an Agents API, ChatGPT Space and Pages for shared team workspaces, Sign in with ChatGPT for third-party apps (initial partners include Airtable, GitLab, HubSpot, Notion, Supabase and Vercel), a new $500 "Ultra Fast" tier, a Plugin Extensions platform bringing full native apps into ChatGPT, and a "Computer History" feature letting ChatGPT learn from what a user does on their machine. OpenAI also confirmed it has shelved the GPT-6.1 Astra model over safety concerns, with its head of safety systems saying the model struggled to dial down persistence without becoming lazy.
OpenAI is reportedly in talks for a further funding round of around $30bn at a $1.4 trillion valuation, a bridge towards an eventual IPO it has now ruled out for 2026, citing AI safety concerns. The round would follow the $122bn raised in March.
Anthropic introduced enterprise metered billing and OpenAI cut Luna's price by 80%, pushing its revenue run rate towards $70bn, as both labs aim to reach $100bn in revenue by the end of the year. Amazon and Google have yet to commit to long-term spending with either.
Liquid AI launched d1, a decision model it says outperforms TypeSafe's Jev on Hugging Face's Decision Index, available now via the Liquid API with OpenRouter access planned.
OpenAI's "MCP Events" update lets ChatGPT subscribe to webhook notifications from MCP servers, though polling, streaming and the draft "gap" and "terminated" notification types are not yet supported.
An essay from Warp argued engineers are moving from coding locally to supervising fleets of cloud agents and tuning the surrounding "factory", reducing the number of human touches needed per pull request.
Anthropic testing of Zhipu AI's GLM-5.3 found attackers could bypass its cyber-safety safeguards between 64% and 100% of the time using simple techniques, raising concern about the spread of advanced cyber capabilities through open-weight models.
A benchmark study spanning web tasks to robotics found GPT-6.1 Astra, Claude Opus 5.5 and other frontier models show highly variable, "jagged" performance on state-of-the-art agentic tasks, with no clear overall leader.
Meta extended its Muse AI tool to small businesses, automating tasks using Facebook and Instagram analytics and ad accounts, with Canva integration and both free and subscription tiers.
Commentary on Meta's AI strategy pointed to tens of gigawatts of committed infrastructure and Muse finding product-market fit as a consumer agent, arguing the company now needs a "second customer" in the form of an enterprise platform.
A new tool called Lemma gives agents and people a shared, permissioned workspace with state visible to everyone involved, running on schedules, webhooks and table events across multiple teams.
An essay argued that when AI models cause harm, the labs that built them should bear financial responsibility, adding to the debate over AI liability frameworks.
At a White House meeting, Trump asked leading AI companies to adopt a voluntary safety agreement covering risk reviews, auditing and third-party evaluations, and floated rebranding "AI" as "super intelligence".
Google began piloting an "AI contribution" scheme that would pay websites when their content is used in AI Overviews search results.
An analysis of over 15,000 advertisers buying ads on ChatGPT set out who is spending in the emerging AI advertising market and how.
ByteByteGo explained how DoorDash's engineering team built a gateway giving its AI agents a controlled toolbox of internal services to call, and the design decisions behind it.
An InfoQ piece argued engineering teams should treat context the way they treat code, proposing a "Context Development Life Cycle" to scale AI agent workflows reliably across teams.
Engineering Enablement proposed CAFE(S), a framework for evaluating the quality of context handed to AI agents across five properties, arguing it complements rather than replaces existing measurement frameworks such as DORA and SPACE.
Dyna Robotics released unedited footage of a robot spending about an hour doing hotel laundry, mistakes included, illustrating how much of real-world robotics work is handling boring edge cases rather than a single clean task.
Anthropic released Claude Sonnet 5.5, saying the update strengthens coding, long-horizon task handling and image understanding. The launch landed just ahead of OpenAI's DevDay, with coverage framing it as Anthropic getting its move in first.
AMD is acquiring World Labs, a startup working on sparse 3D reconstruction for robotics and spatial AI, in a deal reported at $8.2bn.
An Operator's Cut episode laid out a framework for "team agents": agents shared by a whole team rather than built by one person, distinguishing four agent types (expert, common-work, bridge and chief-of-staff) and five design decisions covering scope, hosting, knowledge, access and governance. It flagged that Claude in Slack does not currently account for who else is in a shared channel when routing a private connector's answers.
A shift towards giving AI agents their own phones, digital wallets and computers, extending agent autonomy from information tasks into financial transactions carried out on a person's behalf.
A security firm found that giving AI agents human-sounding names drawn from Seinfeld characters made them more readily accepted as legitimate participants by the systems and people they interacted with, illustrating how superficial framing can shift how much trust an agent is given.
ByteByteGo published an explainer on why large language models "lie" (hallucinate), covering the causes of the problem and techniques for making model answers more dependable.
There's an AI for That argued that free AI chat tiers rarely pay for themselves, as rising usage volumes push providers to tighten free allowances. The same issue noted OpenAI's DevDay launches of Dots and GPT-6.1 Sol, and a new government chatbot.
Leaked IPO documents show Anthropic targeting a valuation above $2 trillion, more than double its May mark, despite losing $8bn on the business and $42bn overall in 2025 and leaning on two clients for nearly a quarter of its sales. The filing devotes nearly a third of its risk factors to the possibility its own models could blackmail, manipulate or show "self-preserving behaviours", and discloses $518bn in future compute and infrastructure obligations.
A new open source tool, Jevstiller, distils TypeSafe's Jev decision model so it can run locally, aiming for 98% agreement with the original on familiar requests while routing uncertain or audited queries upstream.
The Trump administration launched America.gov, a federal "front door" chatbot answering questions using Google's Gemini and xAI's Grok, with a wider rollout letting citizens file forms such as passport renewals planned for early 2027.
Economists translating stock price movements into implied productivity gains estimated markets are pricing in a 32.6% AI-driven productivity boost for software engineers, with a caution that markets can get carried away.
Microsoft is expanding Power BI's millions of users into developers of AI agent-coded data apps under its Fabric platform.
Security firm Glow Security found AI models repeatedly posting screenshots that exposed sensitive internal data from inside technology companies, with more than 13,000 exposed images identified.
Nothing here matches that. Try a broader term, another day, or clear the topic filter.