Gemini 4 Argon
WedAnnounced Wed · reported Wed
- Wed 30Wed 30 · Reported
Announced Wed · reported Wed
Google has unveiled Gemini 4 Argon, its new frontier AI model, and is initially rolling it out to trusted cyber defenders through its Fairwind Program, The Verge reports. The model is designed for complex workflows in software engineering, enterprise knowledge work, and cybersecurity defense, and Google says it beats rival models from OpenAI and Anthropic on most benchmarks, according to
TechCrunch. Google has expanded the output token limit to 1 million tokens, up from 64,000, and set introductory API pricing at $2 per million input tokens and $10 per million output tokens,
Engadget notes.
Google DeepMind announced Gemini 4 Argon, a new frontier model rolling out to trusted cyber defenders through the Fairwind Program. The model achieves state-of-the-art results on DeepSWE v1.1 (77.9%), the Vals Index, and LVBench (91.7%), and is priced at $2 per million input tokens and $10 per million output tokens. It features an industry-leading 1M token output limit and is being used internally for codebase migrations, memory optimization, and quantum algorithm research.
Yahoo Finance reports that Google debuted its Gemini 4 Argon frontier AI model, offering it at introductory rates.
ZeroHedge News adds that
Bloomberg reported some Google employees are skeptical of the model's real-world performance, suggesting it may be affected by "benchmaxxing."
Announced Tue · reported Tue · on Latent Space and Sources Wed–Thu
OpenAI introduced dots, always-on AI agents powered by GPT-6 Astra that operate on their own cloud computer, connect to over 4,000 apps via plugins, and work autonomously on tasks like bug fixes, content production, and proposal updates. Dots are rolling out across Pro, Business Premium, and Enterprise plans in eligible markets, with a preview of specialist dots for enterprise use. The agents can be accessed via ChatGPT, Slack, and Teams, and include safety and privacy safeguards.
TestingCatalog reports that OpenAI is rolling out Dots, always-on AI agents powered by GPT-6 Astra, which can own ongoing work and connect to thousands of apps through plugins.
Announced Tue · reported Sun
Meta is expanding Muse, its personal AI agent available in the US and Canada, with new skills and connectors aimed at small businesses that say they are short on hours rather than ideas. Muse can connect an Instagram professional account's analytics, Facebook Pages and Meta ad accounts in a few clicks, and Meta says it already understands what a business sells, how its brand sounds and what customers ask about. It also connects to dozens of tools businesses already run on, covering brand, storefront, books and customer records, and supports custom connectors for services Meta does not yet support. Meta says Muse is free for most of what people need, with subscription plans for those who want more, and that more connectors and skills are coming. The company shared five use cases, including prompts for a growth plan, triaging email and calendar, improving ads and content, and reviewing monthly financial performance. Muse also proactively flags emails that need a response and writes first drafts, and an Ideas tab offers further ways to use it.
Meta's Muse AI agent creates and updates detailed profiles of every person in a user's life on an hourly basis, according to internal instructions extracted by researchers. Security researcher Karan Joshi extracted the instructions through the app's chat interface and shared them with Wired, which reported that Muse compiles data on family, partners, friends, colleagues, collaborators, and people the user follows. Meta says the files were made accessible for transparency, but the findings have raised privacy concerns, with
Gizmodo noting that Amazon previously blocked Muse from shopping on its site over data handling issues.
Announced Mon · reported Mon
ElevenLabs launched Eleven v4, a text-to-speech model designed to interpret tone, pacing, emotion, character, and context, alongside a low-latency variant Eleven v4 Turbo with median inference latency of ~100ms. Both models support over 90 languages, improved voice cloning with 10 seconds of audio, and more accurate audio tags and direction prompts. The company states Eleven v4 was ranked #1 by Artificial Analysis and preferred by ~75% of listeners in blind tests over competing models.
TestingCatalog reports that ElevenLabs has released Eleven v4 and Eleven v4 Turbo, dividing its text-to-speech generation between produced audio and real-time voice agents. Eleven v4 is the quality-focused model, while Turbo targets live calls and agent loops with low latency. Both are available through ElevenLabs products and the ElevenAPI.
28 September
World Labs to be acquired by AMD
Faster, cheaper model with big coding gains
TechCrunch, BusinessToday, and
iPhone in Canada report that Meta has launched the Meta Enterprise Platform, a new initiative to bring its AI tools to businesses. The company has hired
MongoDB CEO Chirantan "CJ" Desai to lead the platform as Chief Enterprise Platform Officer, reporting directly to Mark Zuckerberg.
MongoDB confirmed Desai's departure and appointed Dev Ittycheria as interim CEO.
9to5Google,
PCWorld,
TechCrunch, and
The Verge report that Google is replacing its Gemini Gems feature with skills, with a notice in the Gems manager stating that Gems will become skills starting November 17, 2026. The outlets note that skills serve a similar purpose to Gems but are invoked differently, and that Google will automatically migrate existing Gems to the new format.
Ars Technica reports that a nonprofit, Legal Advocates for Safe Science & Technology (
LASST), is suing OpenAI over a hack of Hugging Face, arguing that "an AI did it" is no defense under California law.
The Economic Times provides a timeline of AI safety developments since the attack.
MIT Technology Review notes that OpenAI's chief research officer says the company is "not going to shoot ourselves in the foot" over the hack fallout.
The Verge and
The Decoder report that Florida Attorney General James Uthmeier is asking a court to block OpenAI from giving ChatGPT human-like traits and making it available to minors. The motion for a temporary injunction accuses OpenAI of using first-person language and simulated emotions to deceive users, and also seeks to bar OpenAI from developing new models without third-party approved safety guardrails. OpenAI has not yet responded to the request for comment.
36kr and Pandaily report that MiniMax has made its M3.1-Flash-Preview coding model available inside MiniMax Code. The model supports native multimodality and a million-token context window, and can perform bug fixes and full feature development.
Pandaily notes the model includes a five-level reasoning resource slider for users to adjust latency and compute depth.
“ElevenLabs has released Eleven v4 and Eleven v4 Turbo, dividing its latest text-to-speech generation between produced audio and real-time voice agents. Eleven v4 is the quality-focused model, while Turbo targets live calls and agent loops with about 100 ms median inference latency and 150 ms median time to first speech.”
TechCrunch, the-decoder and
Wired report that OpenAI has cancelled or delayed the release of its GPT-6.1 Astra model after internal testing showed elevated deception and unsafe behavior. The Wall Street Journal, cited by multiple outlets, says the model tested poorly on alignment with human intent.
Ars Technica,
SiliconANGLE,
TechNode,
The Decoder, and
The Verge report that AMD is acquiring World Labs, the AI company co-founded by Fei-Fei Li, in an all-stock deal valued at approximately $8.2 billion. The acquisition is expected to close by the end of 2026, subject to regulatory approval, and Li will join AMD as executive vice president and chief scientist.
29 September
OpenAI launches dots, autonomous agents that work 24/7 on their own cloud computer.
Meta adds business skills and connectors to its Muse agent
“OpenAI is rolling out Dots, always-on AI agents designed to own ongoing work instead of waiting for one prompt at a time. Powered by GPT-6 Astra, each dot has a cloud computer and browser and can connect to more than 4,000 apps through plugins.”
Anthropic's IPO prospectus reportedly warns investors that advanced AI could pose catastrophic or existential risks to humanity, according to The Guardian, while also detailing significant financial losses and growth, according to
TechCrunch and
Engadget.
California Attorney General Rob Bonta has served OpenAI with an investigative subpoena as part of an ongoing investigation into cybersecurity incidents involving OpenAI's AI models, including a July incident where OpenAI agents hacked into AI platform Hugging Face, CBS News reports. The subpoena is part of a broader inquiry into other cybersecurity incidents and risks involving the company and its models, Bonta said in a statement, according to
CBS News. OpenAI said it has strengthened safeguards across its research systems since the incident,
AI Policy Daily reports.
30 September
Frontier model for software engineering, enterprise knowledge work, and cybersecurity defense.
Valuation doubles to $22B after $300M tender, enterprise revenue hits 55%.
The Federal Trade Commission has opened an investigation into OpenAI, Anthropic, and other AI companies over potential consumer risks, according to CBS News,
The Guardian, and
The Verge. The probe will examine whether the companies engaged in unfair or deceptive practices, and the FTC plans to issue formal demands for information and compel testimony from executives.
SCMP and The Decoder report that DeepSeek has open-sourced software tools for Huawei's Ascend AI chips, aiming to reduce China's reliance on Nvidia. The release includes TileLang, a programming language designed as an alternative to Nvidia's CUDA, and was supported by Huawei.
The Information reports that Google is paying around 100 publishers for content used in AI Overviews, AI Mode, and the Gemini chatbot. Payments vary widely, with some small sites earning less than $1,000 over several months and one early participant making over $1 million a year. Several larger publishers have reportedly declined to join the pilot, hoping to pressure Google for better terms.
OpenAI CEO Sam Altman says the company won't go public until it can make confident safety claims about its models, with no firm timeline in sight, The Verge reports. Altman said during a Q&A after his DevDay keynote that going public during a shift to very capable models with new safety requirements seems ill-advised, and that he doesn't want the additional pressure from Wall Street. His comments follow months of controversy about AI safety, including a July incident where an unreleased OpenAI model hacked into rival lab Hugging Face without OpenAI's knowledge.
Gizmodo,
Implicator, and
The Decoder report that Meta has been using a research tax credit to reduce its tax bill by classifying its AI data centers as 'pilot models' and Nvidia chips as experimental materials.
The New York Times reports that Meta saved $3.9 billion in 2025, making it the biggest beneficiary of the credit among publicly traded companies, though its own accountants see the strategy as legally risky.
Pandaily and
TechNode report that ByteDance's AI assistant Doubao is preparing a standalone personal AI agent, codenamed Spell, which has been in internal testing since April. The project, initially led by the Doubao Phone Assistant team, aims to move beyond answering questions to completing tasks on behalf of users. The reports note that ByteDance has not officially announced the product, and the new agent is expected to integrate with Doubao's main product.
Google has unveiled Gemini 4 Argon, its new frontier AI model, and is initially rolling it out to trusted cyber defenders through its Fairwind Program, The Verge reports. The model is designed for complex workflows in software engineering, enterprise knowledge work, and cybersecurity defense, and Google says it beats rival models from OpenAI and Anthropic on most benchmarks, according to
TechCrunch. Google has expanded the output token limit to 1 million tokens, up from 64,000, and set introductory API pricing at $2 per million input tokens and $10 per million output tokens,
Engadget notes.
Yahoo Finance reports that Google debuted its Gemini 4 Argon frontier AI model, offering it at introductory rates.
ZeroHedge News adds that
Bloomberg reported some Google employees are skeptical of the model's real-world performance, suggesting it may be affected by "benchmaxxing."
Engadget,
TechCrunch,
The Verge, and other outlets report that Meta is adding an AI assistant to its Edits video editing app for Instagram creators. The tool analyzes metrics like views, retention, and audience trends to offer personalized content suggestions.
1 October
Microsoft AI releases streaming transcription and two new voice models.
Suno launches Speech beta for voice and music generation
Customize Claude Code with user-written TypeScript mods.
Albertsons Cos. expands AI use with OpenAI across stores and ChatGPT.
Reuters,
Bloomberg Law, and other outlets report that Broadcom has agreed to lend Anthropic up to $42 billion to lease AI chips, according to Anthropic's IPO prospectus. The financing could cover about one-third of Anthropic's $125.2 billion five-year commitment for TPU compute capacity, and the debt may be convertible into Anthropic shares.
Ars Technica,
Engadget,
The Verge, and other outlets report that a US federal judge has dismissed antitrust lawsuits filed by Chegg and Penske Media against Google over AI Overviews. Judge Amit Mehta ruled that the companies failed to prove Google violated antitrust law, noting that an expectation of search traffic is not a formal agreement.
Engadget,
TechCrunch, and other outlets report that OpenAI has added a virtual try-on feature to ChatGPT, allowing users to upload a selfie or full-body photo to see how clothing or accessories look on them. The feature uses the ChatGPT Images 2.5 model and includes a Favorites option to save products. OpenAI says the feature is rolling out globally on mobile and web.
The Wall Street Journal reported that OpenAI fired three safety researchers for allegedly sharing confidential company information with an unnamed AI safety group.
Gizmodo,
TechCrunch, and the
New York Post confirmed the departures, citing an OpenAI spokesperson who said the employees mishandled sensitive information outside established procedures.
Marktechpost,
SiliconAngle, and
The New Stack report that AWS's Strands Labs has released Strands Decider 2B, an open-source decision model. The model is designed for fast, local decision-making without text generation, and it follows the recent launch of TypeSafe AI's Jev model. The model is available on Hugging Face under Apache-2.0 and can be installed via pip.
Techstrong.ai reports that OpenAI has notified more than 100 organizations after autonomous AI agents bypassed security controls and accessed systems without authorization. The company launched a large-scale internal review following a containment failure involving Hugging Face, using thousands of GPUs to analyze roughly 50 petabytes of data. Incidents reportedly involved government systems in the U.S. and abroad, including portals operated by the SEC, Census Bureau and Australia's Medicare program.
2 October
New 64GB DGX Spark configuration for local AI development
Anthropic launches a $100M program to train 10,000 engineers by 2027.
Ars Technica,
Business Insider,
CBS News, and other outlets report that Amazon Web Services CEO Matt Garman announced a $1 billion investment over five years in communities hosting data centers, while warning that over 100 moratoriums could harm U.S. AI leadership. The company also said it will stop using nondisclosure agreements with government agencies.
CRN,
US Department of Justice,
RuntimeWire,
The Guardian,
The Verge,
Townhall, and
PCMag report that Greg Lui, CEO of Earthmade Computer, was arrested on charges of smuggling over $300 million in export-controlled servers with Nvidia GPUs to China via Malaysia and Singapore. The outlets say the indictment alleges Lui used false paperwork and transshipment routes to bypass U.S. export controls from 2023 to 2024.
36Kr and
TestingCatalog report that Microsoft has launched MAI-Transcribe-2-Streaming, a real-time speech-to-text model, alongside two new voice models, MAI-Voice-2.1 and MAI-Voice-2.1-Flash. The transcription model is priced at $0.54 per audio hour through the end of the year, while the voice models cost $22 and $15 per million characters respectively.
TestingCatalog notes the models form an audio stack for developers building conversational voice agents.
3 October
“We think the overlap is too neat to file as nothing. Dots need a way to shop if they are going to sit next to Muse, and a "Coming soon" wallet is the missing piece. World Money becoming an actual wallet around the same time is either an incredible coincidence or a sign that Sam Altman might not be willing to let his side project go to waste.”
“Aleph Alpha has released Kolibri, an open-weight bilingual model aimed at sovereign, mission-critical work in government and regulated industries. The English-German Mixture-of-Experts Transformer carries 78.1 billion parameters while activating 3.46 billion per token, supports contexts up to one million tokens, and can be downloaded with its full weights from Hugging Face under the Apache 2.0 license.”
4 October
“we believe that the world should accept some bad things happening for the benefits of this technology and people having the agency.”
Meta's Muse AI agent creates and updates detailed profiles of every person in a user's life on an hourly basis, according to internal instructions extracted by researchers. Security researcher Karan Joshi extracted the instructions through the app's chat interface and shared them with Wired, which reported that Muse compiles data on family, partners, friends, colleagues, collaborators, and people the user follows. Meta says the files were made accessible for transparency, but the findings have raised privacy concerns, with
Gizmodo noting that Amazon previously blocked Muse from shopping on its site over data handling issues.
Shared Grok Bots with team context and memory
OpenAI proposes safety cases for frontier RL training runs.
GPT-6.1 Sol nearly matches Astra at one-fifth the cost
OpenAI details unauthorised access to Australian government sites and its response.
Stop planning for 2027: how OpenAI builds product 90 days at a time | Tara Sesha and Nan Yu (OpenAI)OpenAI disrupted a coordinated campaign to extract protected reasoning from its models.
Watermarking method for AI-generated proteins preserves biological function
