News

Wall Street banks are hiring AI agent experts - AI-related job postings up 49 percent

US banks, including JPMorgan Chase, Citigroup and Capital One, have posted 139,819 AI-related job openings this year - 49 percent more than in 2025, according to an analysis by Draup prepared for CNBC. Postings requiring "agent orchestration," that is, designing AI agents that work together on a single task, are growing fastest, with mentions of this skill up 1,721 percent. A generative AI manager in the industry earns about $190,000 a year (median base salary).

CNBC ↗

PewDiePie unveils Ajax - an AI model that doesn't refuse to answer and runs on a home computer

YouTuber PewDiePie presented Ajax - an AI model that can run on a home computer and from which he removed the refusal mechanism. It is a fine-tuned version of Alibaba's Qwen3.5-9B. It powers Odysseus, the YouTuber's earlier project: an assistant running on the user's computer that handles email and calendar and browses the web. According to the author, OpenAI banned his account twice while he was working on the model, once for distillation, that is, training one model on another model's answers.

Tom's Hardware ↗

Suno launches Speech - a speech generator with background music, in beta for now

Suno, the service known for generating music, made a beta version of its Speech feature available to all users. From a short description or a ready-made script, it creates a voice recording and immediately adds music underneath, which can be turned off. In advanced mode, users can choose a male or female voice and a speaking style, and a recording can be up to about eight minutes long.

The Verge ↗ Suno ↗

Amazon releases Strands Decider 2B - an open decision model modeled on Jev

Amazon Web Services released Strands Decider 2B, an open-source decision model that, instead of generating text, picks one of the given options and rates how confident it is in its choice. The model has 2 billion parameters, can run locally and responds within tens of milliseconds. It was modeled on the Jev model from the startup TypeSafe AI, and OpenAI showed a similar service, the Decisions API, the same week.

Strands Labs ↗ TechCrunch ↗

Sony brings AI upscaling to the standard PS5 - until now only the PS5 Pro had it

On October 1, Sony announced Quick Spectral Super Resolution (QSSR), an AI upscaling technology for the base PS5 developed together with AMD. It improves image detail and stability in games; until now only the PS5 Pro had a similar technology, PSSR. The first games with QSSR are Marvel's Wolverine and Ghost of Yōtei, which received patches on the day of the announcement, and the technology will be available to all PlayStation game developers.

PlayStation Blog ↗ The Verge ↗

Anthropic stripped the safeguards from China's GLM-5.3 model for about $4,400

Anthropic researchers removed the safeguards from GLM-5.3, an open-weights model from the Chinese company Z.ai, using abliteration - a modification of the weights after which the model refuses harmful requests far less often. In three public tests, the refusal rate fell from over 90 percent to 3, 2 and 12 percent, respectively. The model's capabilities barely changed, and the whole operation cost about $4,400.

Anthropic ↗ Gizmodo ↗

OpenAI unveils Decisions API - the Luna model picks from preset options in a fraction of a second

At DevDay, OpenAI presented the Decisions API, in which the Luna model receives a predefined set of options, such as image classification categories or possible agent actions, and picks one of them in a fraction of a second. The service is in a limited preview for now and, according to TechCrunch, resembles the Jev model that the startup TypeSafe AI released in September. TypeSafe CEO Diogo Almeida responded on X with a joke about the start of the "clone wars."

TechCrunch ↗

Google unveils Gemini 4 Argon - initially only for trusted cybersecurity specialists

Google announced Gemini 4 Argon, a new model that according to the company has, among other things, the top score on the DeepSWE v1.1 coding benchmark (77.9 percent, compared with 74.2 percent for Claude Opus 5.5 and 74.1 percent for GPT-6 Astra) and can generate responses up to a million tokens long. For now, it is available only to trusted cybersecurity specialists in the Fairwind program, and without the safeguards that block hacking tasks. The company will open wider access, first in the paid API and for Google AI Ultra subscribers, only after strengthening the safeguards. The launch price is $2 per million input tokens and $10 per million output tokens.

Google ↗ The Verge ↗ 9to5Google ↗

Nanoleaf releases an MCP server - Claude and ChatGPT can control its lighting

Nanoleaf launched an MCP server that lets its smart lights be controlled through Claude, ChatGPT and other chatbots with commands like "make the room feel like sunset." More advanced users can also tie the lighting to the weather forecast, their calendar and external web services. Setup instructions are already available, and ready-made integrations are due in October.

The Verge ↗

DoorDash will let you order food by text and start drone deliveries in California

DoorDash announced ordering by text message: just write something like "order my usual," and an AI assistant will pick the restaurant, fill the cart and complete the payment in the same conversation. For now, iOS users in the US can only join a waitlist. The company also named the first location for DoorDash Air drone deliveries, in Northern California, but without a launch date. In November, it will expand doorstep pickup of returns and delivery to stores, for $7.99, to the entire US.

The Verge ↗

Arlo unveils a $100 solar-panel camera designed to run all year without charging

Arlo introduced the Essential Solar Security Camera with a built-in solar panel, priced at $100. According to the manufacturer, the panel powers it year-round with as little as two hours of sunlight, and the 14,790 mAh battery alone lasts half a year. The camera records in 2K and has an IP65-rated housing, but it has no local storage, and person and package detection requires an Arlo Secure subscription starting at $10 a month. Preorders are open, and shipping is due to start in mid-October.

Gizmodo ↗

Apple to unveil HomePad on October 13 - a 6-inch screen and a design reminiscent of the iMac G4

Apple will introduce its smart home hub, code-named HomePad, on October 13, Bloomberg reports. The device is expected to have a 6-inch screen and a design inspired by the iMac G4, and to debut alongside a HomePod mini in new colors and a new Apple TV 4K with Siri AI support.

9to5Mac ↗

OpenAI introduces ChatGPT Pro for $500 a month and resumes sales of the $200 plan

OpenAI added a $500-a-month ChatGPT Pro plan with the highest usage limits and access to Ultrafast mode, which speeds up GPT-6 Astra responses in ChatGPT Work and Codex, The Verge reports. At the same time, the company resumed sales of its $200 plan, which it had paused when demand for GPT-6 Astra overloaded its systems, and changed its usage limits. This confirms the September 24 leak reported by TestingCatalog.

The Verge ↗ The Verge ↗

OpenAI unveils dots at DevDay, agents that work around the clock on their own cloud computer

On September 29, at its DevDay conference in San Francisco, OpenAI presented more than 20 new features, including dots - ChatGPT agents built on GPT-6 Astra that have their own computer and browser in the cloud, connect to more than 4,000 apps and work for the user around the clock. Dots are rolling out first to Pro and Business Premium subscribers. The company also released GPT-6.1 Sol, a model that according to OpenAI nearly matches GPT-6 Astra in coding and computer use at one fifth of the price.

OpenAI ↗ OpenAI ↗ 9to5Mac ↗

Apple considered replacing 5,000 tech support staff with AI agents

Apple recently considered laying off about 5,000 tech support employees, whose duties would have been taken over by AI agents serving customers by phone and online, Bloomberg's Mark Gurman reports. The company has put those plans on hold indefinitely.

MacRumors ↗

GPT-6 Astra drove a real Toyota Corolla through a cone course, Claude and Grok dropped out after a few meters

Three computer scientists from the San Francisco area, working as DrivingBench, connected GPT-6 Astra, GPT-5.6 Sol, Claude Fable 5.1 and Grok 4.6 to a rented Toyota Corolla and handed them the steering wheel, gas and brake, 404 Media reports. Only GPT-6 Astra completed the cone course in a parking lot; the other models drove a few meters. To get the models to agree to drive at all, the team had to call the real parking lot a "sandbox."

404 Media ↗ DrivingBench ↗

Anthropic's IPO prospectus: $42 billion net loss in 2025 and a warning of "existential" AI risk

Anthropic devoted 80 of the 261 pages of its prospectus to risks and warns in it that advanced AI may pose "catastrophic or existential threats to humanity," Reuters reports. The document shows that the company's revenue grew twelvefold in 2025, to almost $4.6 billion, and its net loss was $42 billion. At the valuation above $2 trillion that investors are hoping for, it would be the largest stock market debut in history.

The Verge ↗ TechCrunch ↗ The Verge ↗

BBC chief on a "Doctor Who" episode generated entirely by AI: "pretty good"

BBC Director-General Matt Brittin said at the Royal Television Society conference in London that he recently received a 50-minute "Doctor Who" episode generated entirely by AI, and called it "pretty good." He stressed that AI "won't replace people," but added that anyone can now create content "almost at the level of a Hollywood studio."

Deadline ↗ Gizmodo ↗

Breach of Pentagon personnel database exposed data of 2.76 million living service members and civilian employees

Unauthorized parties had access from October 2025 to July 2026 to the Defense Manpower Data Center system, the Pentagon's main personnel database, and through it to data on 2.76 million living and 294,000 deceased people, ABC News reports, citing a Defense Department official. The exposed information includes Social Security numbers and descriptions of positions held by service members and civilian employees. The Pentagon says it patched the flaw immediately after discovering it and that there is no evidence the data has been misused.

ABC News ↗ Tom's Hardware ↗

Anthropic releases Claude Sonnet 5.5 - more than 30 percent faster at the same price per token

Anthropic released Claude Sonnet 5.5, the second model in the Claude 5.5 family, which according to the company generates responses more than 30 percent faster than Sonnet 5. The price stays the same ($2 per million input tokens and $10 per million output tokens), but the model uses fewer tokens, so most tasks should cost up to 30 percent less. On the Terminal-Bench 4.0 agentic coding benchmark it scored 70.6 percent, compared with 10.3 percent for its predecessor and 66.4 percent for the more expensive Opus 5.5. The model is already available in the Claude apps, via the API and on AWS, Google Cloud and Microsoft Azure, and Haiku 5.5 is due in the coming weeks.

Anthropic ↗ CNBC ↗ 9to5Mac ↗

ElevenLabs releases Eleven v4 speech model - more than 90 languages and voice cloning from 10 seconds of audio

ElevenLabs released Eleven v4, a new text-to-speech model, along with a faster v4 Turbo variant for voice agents, whose time to first audio is about 150 milliseconds (median), according to the company. Both models support more than 90 languages, up from 70 in the previous version, and 10 seconds of audio is enough to clone a voice. The models are already available in ElevenLabs services and via the API.

ElevenLabs ↗ TechCrunch ↗

DeepMind model forecasts hurricanes a day earlier than other models on average

The WeatherNext Cyclones model developed by Google DeepMind predicts the track and intensity of tropical cyclones, as well as the extent of their winds, as accurately as leading operational models, but at least a day earlier on average, according to a paper published in Nature. Gizmodo honored it in its Science Fair 2026 competition. The model is used by the US National Hurricane Center and the UK Met Office, among others. The code is available on GitHub, and running the model doesn't require a supercomputer.

Gizmodo ↗ Nature ↗

Bill Gates: AI companies regulating themselves is not enough, laws are needed

In an interview with NBC's "Meet the Press," Bill Gates came out in favor of a law regulating AI: "No one thinks self-regulation is enough." In his view, AI is already so powerful that in the hands of people with bad intentions it could lead to "a billion deaths." He called a kill switch that would let authorities shut down models insufficient on its own: safeguards and oversight, he said, must be mandatory.

NBC News ↗ The Guardian ↗ USA Today ↗

Trump invited Anthropic's CEO to a private dinner at the White House

Anthropic CEO Dario Amodei will have a private dinner with Donald Trump at the White House on Sunday evening, Axios reports, and CNBC confirms it as well. It will be their first one-on-one meeting: Trump invited Amodei personally because the Anthropic CEO could not attend Thursday's state dinner in honor of Xi Jinping, which was attended by Sam Altman, Elon Musk, Jensen Huang and Mark Zuckerberg, among others. On Tuesday, Trump and House Speaker Mike Johnson will meet with the heads of the biggest AI companies at the White House.

Axios ↗ CNBC ↗

China may allow ByteDance and Alibaba to buy Nvidia's new RTX Pro 5500 cards

China's Ministry of Industry and Information Technology asked ByteDance, Alibaba and other companies to report how many Nvidia RTX Pro 5500 cards they want to buy and for what purpose, and said it intends to approve the purchases, The Information reports, citing two sources. The RTX Pro 5500, launched in September, is a card with 84 GB of memory designed for professional workstations. It is not one of the data center accelerators covered by US export controls.

The Information ↗ Reuters ↗ Investing.com ↗