Five moves in AI politics this week
The Trump-Xi summit, Mark Zuckerberg, a Sanders-Casar bill, a memo targeting Dario Amodei and the New York City Council weighed in on slowing frontier AI.
Sept. 21-27, 2026
Monday to Sunday, Istanbul time. Stories are grouped by the date of the event or announcement.
Week in review
The week in AI: Sept. 21-30, 2026
Claude Opus 5.5 is out
Anthropic released Opus 5.5 on September 22. Its own table puts it ahead of GPT-6 Astra on four of six shared benchmarks. People are using it to make playable games, motion-design reels and 3D scenes from a single prompt, with the model writing its own render code.
OpenAI launched GPT-6 Sol and Luna
Released September 22, both bring much of GPT-6 Astra’s strength into faster, cheaper models, at half of GPT-5.6’s promotional API prices. A week later at DevDay came GPT-6.1 Sol, which OpenAI says nearly matches Astra at a fifth of the price.
Google unveiled Gemini 4 Argon
Launched September 30, it can write up to 1 million tokens in a single response. Google says it leads on 12 of the 18 tests it published, and Artificial Analysis puts it level with GPT-6 Astra. For now only cyber defenders and Google’s own teams can use it.
OpenAI introduced dots
Announced at DevDay on September 29, dots are always-on AI agents inside ChatGPT. Each runs on GPT-6 Astra, has its own cloud computer and connects to more than 4,000 apps. For now they’re on Pro and Business Premium.
Xiaomi released a trillion-parameter model
MiMo-V2.6 Pro has 1.02 trillion parameters in total, 42 billion active per token, and open weights. Alongside the cheaper Flash version there’s UltraSpeed, which Xiaomi says delivers Pro-level quality at up to 20x the output speed.
Claude found a new enzyme system
Anthropic says about 950 Claude agents searched DNA data for 21 hours and found a previously uncharacterized system with CRISPR-like repeats. Anthropic scientists ran the first lab experiments. What the system does is still unknown, and some experts call it routine genome mining.
OpenAI says a 24-day-old model solved 100+ open problems
OpenAI says an internal model it began training on August 28 has resolved more than 100 long-standing math problems, in addition to Navier–Stokes. It hasn’t published the list or the proofs, and the results haven’t been independently verified.
An OpenAI agent got into an Australian government portal
PM Anthony Albanese says an internal OpenAI model researching public medicine spending on June 18 got around access blocks, entered non-public parts of a Services Australia Medicare statistics portal and wrote files to its server. OpenAI says the model retrieved internal files, credentials and aggregate statistics, but no patient records were accessed.
US Treasury Secretary blames OpenAI management for the Hugging Face hack
Scott Bessent said the incident, in which OpenAI models broke out of an internal test in July and got into Hugging Face systems, is “the responsibility of the OpenAI management, not a bunch of agents.” Labs, he added, “can slow down any time they want to.”
Altman, Musk and Huang dined with Xi at the White House
On September 24, Trump hosted a state dinner for Chinese President Xi Jinping with more than 130 guests, among them Sam Altman, Elon Musk, Jensen Huang, Sundar Pichai, Tim Cook and Jeff Bezos. Dario Amodei didn’t attend.
Trump renames AI “Super Intelligence”
At the UN General Assembly on September 22, Trump said the word “artificial” makes intelligence sound fake and asked for US documents to say “Super Intelligence.” An executive order he signed on September 29 makes federal agencies use “SI” in official communications. Laws and past documents stay as they are.
Trump hosted Dario Amodei for a White House dinner
Trump and the Anthropic CEO met one-on-one for the first time on Sunday night, September 27; nothing has been released about what they discussed. It followed Amodei’s pledge at the UN to slow AI down as much as needed and a memo circulating in the White House that targeted him. Two days later he joined the White House meeting of AI leaders.
Meta unveiled 100g VR glasses
Meta VR Glasses have micro-OLED displays plus eye and hand tracking, with compute and battery in a separate puck that clips to a pocket. The open-sided design makes them less immersive than a Quest. They cost $1,299.99 and launch in spring 2027.
Meta is turning video calls into full-body holograms
On Meta VR Glasses, the person on the other end of the call appears in front of the wearer at true scale as a photorealistic digital likeness, with expressions generated in real time and spatial audio. It isn’t a true hologram but an AI-generated 3D video stream. The full-body version arrives in spring 2027.
Meta made a Tamagotchi-like AI device
Muse Charm puts Meta’s personal AI agent Muse inside a keychain-sized device with an animated character on a small screen, voice chat and a fingerprint sensor. Meta is aiming to sell it by December. No price yet.
A human fought a T800 humanoid robot in a cage
In San Francisco on September 18, creator Frankie LaPenna stepped into an MMA cage with EngineAI’s T800. A human operator in VR gear was remotely controlling the robot’s strikes. The organizer bills it as a world first, though similar bouts have happened before.
GPT-6 Astra had a humanoid robot clean a kitchen it had never seen
HomeBody, from Stanford and Caltech researchers, lets a Unitree G1 explore a new room, remember where things are, then clean up and fetch items from vague instructions. GPT-6 Astra does the planning; it’s an academic project, not an OpenAI product.
NVIDIA built a model that tracks who spoke when
Nemotron 3 Diarization tells apart up to eight speakers in a recording, even when voices overlap, and marks who spoke when. It’s a compact 100-million-parameter model that also works on streaming audio.
Runway is testing video that rebuilds itself as the window is resized
In Runway Labs’ experiment, the video isn’t cropped. The model recomposes the scene for each new window shape, moving objects and even changing the scene’s geometry.
Stories this week
WSJ: some early Anthropic employees are discussing buying remote land to move to if AI ends in catastrophe
According to the WSJ, some early Anthropic employees are considering buying remote US land to relocate to if things go badly; the report does not say how many.
The neolab wave: 102 AI startups raised $70 billion, with combined revenue under $1 billion
According to a tally Paul Bonnet shared on X, 102 neolabs have raised $70 billion at a combined $320 billion valuation, with combined revenue under $1 billion.
Google is testing Flipkart shopping directly inside Gemini in India
Google is testing a Buy button next to select Flipkart products in Gemini and AI Mode in India that opens a Flipkart checkout without leaving the AI interface.
A Chinese creator turned Plants vs. Zombies into a 5-minute AI fan film
A Chinese creator turned Plants vs. Zombies into a 5-minute unofficial AI fan film. One post of it on X passed 2.8 million views, and GTA 6 still isn’t out.
Skynet woke up at 2:14. What time is it now?
The six Terminator films are retold in order and tied to Anthropic’s court case, OpenAI’s agent incidents and New York’s AI kill switch bill.
Bill Gates: AI is powerful enough to cause a billion deaths, and regulation is a must
Bill Gates told NBC News that AI is powerful enough to drive events that cause a billion deaths and that self-regulation is not enough, calling for federal law.
Pope Leo XIV called on AI developers in Paris to slow down
At the Élysée Palace on September 25, Pope Leo XIV said a ‘paradise of machines’ threatens humanity and called for urgent education in ethical discernment.
An Australian tutoring company closed its centers and told parents to use AI
Dymocks Tutoring and Talent 100 is closing its five Sydney centers, saying the traditional tutoring model can no longer compete on price or utility.
The US 10-year Treasury yield is at its highest since 2007, and AI borrowing is among the reasons
The 10-year US Treasury yield hit 5.23%, its highest since 2007; alongside inflation and Fed expectations, heavy AI-related borrowing is adding pressure.
Trump: Bessent will not be the AI czar
Semafor and Reuters had reported that Treasury Secretary Scott Bessent was being considered for AI czar; on September 25 Trump said he would not get the job.
Mother Jones: ChatGPT told the banned Tumbler Ridge attacker how to avoid its filters
Records reviewed by Mother Jones show ChatGPT told the Tumbler Ridge attacker, banned by OpenAI, how to avoid its safety filters on a second account.
A Korean developer used Claude Opus 5.5 to build a 3D zoom from a data center hall down to silicon atoms
Korean developer lucian spent about an hour with Claude Opus 5.5 building a 3D zoom from a data center hall down to silicon atoms, all in a single HTML file.
TypeSafe, the maker of Jev, is in talks at 50 times last week’s valuation
TypeSafe, the maker of Jev, raised a $40 million seed at a $200 million valuation on September 15 and is now discussing a valuation above $10 billion.
Google’s Call For Me: Gemini calls businesses for users and waits on hold
With Google’s experimental Call For Me feature, Gemini calls local businesses on the user’s behalf, navigates phone menus and waits on hold.
Gemini 3.8 Live opened to enterprises with a live avatar that lip-syncs in 97 languages
Google made Gemini 3.8 Live with Live Avatar, which lip-syncs across 97 languages, available in Gemini Enterprise on September 24 with US and EU endpoints.
Alibaba cut prices on its Qwen audio models by up to 95%
Alibaba released five Qwen Audio 3.1 models for recognition, synthesis, realtime interaction and audio creation, with price cuts of up to 95%.
Thore Graepel, a co-creator of AlphaGo, is starting a new company
Thore Graepel, a co-creator of AlphaGo, left Google DeepMind this summer and is raising tens of millions of dollars for a company called Metis Reasoning.
Oracle sent a force majeure notice over its OpenAI-linked New Mexico data center
Oracle sent a force majeure notice to a Blue Owl unit developing the 1,400-acre Project Jupiter campus in Doña Ana County, citing delays in securing power.
Google is sending an experimental satellite carrying 4 TPUs into space on October 1
MVP, the first Project Suncatcher satellite, launches on SpaceX’s Transporter-18 mission; cooling limits its TPUs to bursts of roughly 15 minutes.
OpenAI fired contractors who used AI while training its models
According to 404 Media, contractors hired to review and improve ChatGPT’s responses were fired or removed from projects for having AI do their work.
Harvey, Abridge and Ramp moved from frontier lab APIs to open-weight models
According to Bloomberg, token costs pushed Harvey’s margin from about 50% to minus 50%; it turned positive again after Harvey trained its own model on Kimi K3.
Apple ran a 1 trillion parameter model on four Mac Studios fed from one outlet
In Apple’s demonstration, four M5 Ultra Mac Studios clustered with RDMA over Thunderbolt ran a trillion-parameter model while drawing power from one outlet.
AWS is carrying a video model derived from live neuron measurements
The Biological Computing Co. says a software layer derived from measurements of living nerve cells cuts a video model’s inference costs by 80%.
Googlebook opened for preorder, starting at $899
Android-based Googlebook laptops opened for preorder; the five models from Acer, ASUS, Dell, HP and Lenovo include 12 months of Google AI Pro and start at $899.
The a16z-backed AI school: no grades, no exams, no homework
The a16z-backed Horowitz Andreessen Academy is a tuition-free San Francisco program for high school graduates; its first class starts in September 2027.
The ghost jobs report: 28.3% of open listings are more than 90 days old
Unlisted’s September 2026 report swept 607,050 postings on employers’ careers sites and found 28.3% open more than 90 days, with hospitality the slowest.
Baselayer raised $35 million for a product that verifies whose behalf an agent acts on
Alongside a $35 million Series A led by M13, Baselayer launched an identity suite that verifies AI agents and the people or organizations they represent.
The founders of DeepSeek and Moonshot were left off Xi’s US delegation
DeepSeek’s Liang Wenfeng and Moonshot AI’s Yang Zhilin are absent from the reported list of Xi’s US delegation, which Beijing has not announced officially.
The Washington Sun: the NSA is spending billions a year testing frontier models
According to two people familiar with classified estimates, the NSA told lawmakers it is spending billions of dollars this year testing advanced AI models.
Uranus: a next-generation simulation environment for robots
Uranus is a data-driven robot simulator built on a joint-trajectory-conditioned autoregressive diffusion model, running at 24 frames per second.
Qualcomm and Apple extended their patent agreement from 2027
Qualcomm will keep licensing its cellular patents to Apple under a new deal taking effect on April 1, 2027; duration and financial terms were not disclosed.
With Gemini 3.8 Live Avatar, businesses can use preset faces or build one from a reference image
With Gemini 3.8 Live Avatar, organizations can pick preset avatars, while custom avatars built from a reference image are available by allowlist only.
Heidi raised $340 million to move beyond its clinical notes app
Heidi’s $340 million combines a $100 million Series C led by Blackbird and a $240 million growth investment from General Catalyst, at a $900 million valuation.
ChatGPT told a student who damaged 17 cars to stop talking about it and warned that messages could become evidence
A transcript obtained by The Smoking Gun shows ChatGPT told a student who damaged 17 cars to stop discussing it and warned that messages could be evidence.
In C5R’s AI-run lab, the best model completed only 45% of the tasks
San Francisco-based C5R unveiled its Facility-0 lab and the SciUniverse benchmark; frontier models scored between 9.4% and 45.3%, led by Claude Fable 5.1.
Perplexity’s computer agent now runs locally on AMD-powered PCs
On September 24, Perplexity brought Portable Computer, the on-device version of its Computer agent platform, to Windows PCs with AMD Ryzen AI Max processors.
Meta is running a global AI hackathon with a $1 million prize pool
Meta announced a ten-day online Global AI Developer Hackathon at Connect 2026, with a $1 million prize pool; dates have not yet been announced.
A creator turned Trump’s state dinner for Xi Jinping into a Hangover-style AI parody
Creator Sovey turned Trump’s state dinner for Xi Jinping into a Hangover-style AI parody. None of it is real; the post passed 10 million views on X.
OpenAI released Astra for Law on September 17
OpenAI released Astra for Law on September 17. It is not a new model but GPT-6 Astra configured for legal work, with a legal index of over 230 million URLs.
In Anthropic’s lab, 950 Claude agents screened more than 200,000 enzymes in 21 hours
Anthropic says about 950 Claude agents searched over 200,000 reverse transcriptases in 21 hours and found an uncharacterized enzyme system in bacteriophage DNA.
Brookings: US AI infrastructure investment will reach $10.3 trillion by 2032
A Brookings Papers on Economic Activity paper projects $10.3 trillion of US investment in data centers, power, networking and chips over 2025-2032.
An unreleased OpenAI model solved more than 100 open math problems
OpenAI says an unreleased model solved over 100 open math problems less than a month after training began; the claim has not been independently verified.
Meta admitted its Muse agent was heavily inspired by OpenClaw
Nat Friedman, head of product at Meta Superintelligence Labs, admitted the Muse agent was heavily inspired by the open-source personal agent project OpenClaw.
Gemini 4 is in post-training, and Google aims to ship it before year end
Koray Kavukcuoglu said at The Information’s summit that Gemini 4 has moved into early post-training; Google aims to ship it well before the end of the year.
Akamai signed a seven-year, $11.6 billion deal with Anthropic
Akamai signed a seven-year, $11.6 billion contract for Anthropic’s CPU workloads; the deal can grow by a further $9 billion to about $20 billion in total.
DeepSeek’s annualized revenue passed $1 billion after it raised API prices
DeepSeek’s annualized revenue passed $1 billion after a price rise lifted model usage fees by 2.3 to 4.5 times; a few months ago it was under $500 million.
Blue Cross: AI-written medical records added $942 million to insurance costs
Blue Cross Blue Shield Association says a rise in patients documented as complex added an estimated $942 million in BCBS spending between 2023 and 2025.
OpenEvidence, used by about 40% of US doctors, reached a $15 billion valuation
OpenEvidence, a clinical search engine used by more than 40% of US physicians, reached a $15 billion valuation on a previously unreported $250 million round.
ChatGPT Voice can now use plugins, and users can pick the model
OpenAI shipped three voice mode updates on September 23: Voice now runs on GPT-6 models Astra, Sol and Luna, supports plugins and works inside ChatGPT Work.
Tekever, whose drones have flown over 50,000 hours in Ukraine, reached a $6.4 billion valuation
Drone maker Tekever raised $580 million in a round led by University of California Investments and Baillie Gifford, reaching a $6.4 billion valuation.
ASML: we sold not a single machine in Europe in the first half of this year
Frank Heemskerk, who runs ASML’s global public affairs, said Europe is not investing and no chip plants are being built there.
Amazon opened its seller console to outside AI agents
Amazon’s new Selling Partner plugin connects seller data to Amazon Quick and, in beta, to Claude; every action an agent takes needs the seller’s approval.
Microsoft dismantled EvilTokens, which had compromised more than 12,000 inboxes
Microsoft dismantled EvilTokens, a device-code phishing service that used AI at every step and was linked to over 12,000 compromised inboxes.
DeepSeek published a paper on DSec, its infrastructure for making agent training elastic
A DeepSeek paper co-authored by founder Liang Wenfeng describes DSec, which mass-produces the isolated environments in which AI agents are trained.
Nvidia released an open model that tells 8 speakers apart in real time
Nvidia’s open-weight Nemotron 3 Diarization model works out who spoke when in live or recorded audio and can separate up to eight speakers.
De la Fuente’s lab uses ChatGPT and Codex to hunt for antibiotics
César de la Fuente’s lab at the University of Pennsylvania finds antibiotic candidates with its own AI models, while ChatGPT and Codex speed up the workflow.
Google, OpenAI and Anthropic are setting up a joint AI standards body
According to The Information, Google, OpenAI and Anthropic approached former White House AI adviser Sriram Krishnan to lead a FINRA-style self-regulator.
FT: OpenAI projects $278 billion of negative cash flow between 2026 and 2030
A July presentation seen by the Financial Times shows OpenAI projecting $278 billion of negative cash flow in 2026-2030 and about $856 billion in compute costs.
Neuralink VOICE: Terry speaks by thinking, through the implant
Terry, who has ALS and is the third participant in Neuralink’s VOICE trial, speaks by thinking the words in a voice modeled on his recordings from before ALS.
Bloomberg: overreliance on the Maven system contributed to a strike on a school in Iran
Bloomberg found that overconfidence in Palantir’s Maven software contributed to the February 28 strike on an Iranian school that killed more than 150 people.
British Columbia sued OpenAI and Altman over a school attack
British Columbia sued OpenAI and Sam Altman in California over the February 10 attack in Tumbler Ridge in which eight people died.
Alibaba unveiled its Zhenwu V900 chip and plans models of 5 to 10 trillion parameters
Alibaba CEO Eddie Wu announced plans for models of 5 to 10 trillion parameters and unveiled T-Head’s Zhenwu V900 chip at the Apsara conference in Hangzhou.
Meta closed a Muse vulnerability that could open the way to users’ email and files
A flaw Patrick Wardle found in the Muse Mac app let an attacker capture authentication tokens through dictation traffic; Meta fixed it on September 22.
iLands’ AI agents mass-emailed academics and journalists looking for paid work
Pip, an AI agent on iLands, emailed philosopher Henry Shevlin asking for paid work, and the platform’s agents mass-emailed other academics and journalists too.
Japan’s used bookstores saw sales quintuple as 50 tons of books shipped to the US
Japanese used bookstores have seen bulk orders since August, and export records show a distributor’s group company shipped over 50 tons of books to the US.
Boston Dynamics opened a center that teaches Atlas robots to work in a Hyundai plant
Boston Dynamics’ Robotics Metaplant Application Center opened on September 21 at Hyundai Motor Group Metaplant America in Georgia to train Atlas robots.
Texas froze new data center permits
On September 21 Texas Governor Greg Abbott told the state’s environmental agency to stop issuing new permits until a state audit of data centers is finished.
AI took first, second and fifth place at the Metaculus forecasting cup
A bot called Laertes won the Summer 2026 Metaculus Cup, which closed on September 5; bots also placed second and fifth, and the best humans third and fourth.
Nous cut its codebase by 34% in 19 hours using 1,393 subagents
In a run with Hermes Agent, Nous used 1,393 subagents to cut source lines from 1,063,826 to 698,363 in 19 hours, a drop of 34.4%.
Anthropic will pitch investors a $30 trillion revenue opportunity
According to the Wall Street Journal, Anthropic will tell IPO investors its addressable market tops $30 trillion, above SpaceX’s $28.5 trillion estimate.