AI media releases indexed by MediaRelease.co from verified government and company sources. Listings include a concise MediaRelease.co summary and access to the full release and original source.
Only media releases directly from companies themselves or government departments are referenced on these pages.
Baseten reported that Claude Code with Fable 5, using MetaInfer skills on a single NVIDIA B200, built a Qwen 3.6 inference engine that beat vLLM by up to 90% on decode speed and 2.3× on time-to-first-token. The blog post describes a skills-only agentic workflow for custom open-model serving engines without model post-training. Baseten published the results on 2 October 2026.
Voyage AI introduced rerank-3 and rerank-3-lite, the next generation of its text rerankers for search and retrieval-augmented generation. Rerankers reorder embedding-retrieved candidates so the most relevant documents rank first; the new models improve on rerank-2.5 on quality and latency. Both variants are available through the Voyage API, with rerank-3-lite aimed at lower-cost deployments.
Black Forest Labs released FLUX 3 Action, a 7-billion-parameter open-weights world action model that predicts agent actions from visual observations for robots, simulators, and games. The model extends the company's FLUX visual stack into action prediction for physical and simulated control. Black Forest Labs published the post with open weights for the 7B checkpoint.
ServiceNow CoreAI published AutoSynthData on Hugging Face, a pipeline that turns enterprise agent capability gaps into validated training tasks. The method uses a target model’s failures and a stronger teacher’s successes to choose what to generate next, then checks new tasks against the environment. The authors illustrate the approach with EnterpriseOps Gym; the Hugging Face blog post is dated Oct. 2, 2026.
NVIDIA said a 64GB unified-memory configuration of DGX Spark will be available from Acer, ASUS, Dell, Gigabyte, HP and MSI starting Friday, Oct. 23, 2026. The personal AI system keeps the GB10 Grace Blackwell Superchip, DGX OS and NVIDIA AI software stack, supporting local agents and models up to about 100 billion parameters on device. Two units can be clustered with NVIDIA Sync Cluster Assistant; NVIDIA reported up to 1.7x performance versus one system on a Qwen 3.8 27B test.
Google’s Project Suncatcher prototype satellite, built with Planet, launched into orbit aboard SpaceX’s Transporter-18 rideshare mission. The team confirmed contact and that the satellite is operating as expected. The flight is a first step toward exploring scalable machine learning infrastructure in space and will gather in-orbit data on how TPUs handle stress, radiation, and thermal extremes. A peer-reviewed paper on the research is available in Joule.
OpenAI published a customer story on The Den Family Social, a Denver parent-focused social club using ChatGPT Work while opening a second location. The leadership team reports saving 10–15 hours a week, cutting grant applications from three days to two hours and liquor-license materials from four days to three hours, with Gmail, Slack, and Google Drive plugins. Founder Chandler Lipe’s team also uses ChatGPT for partnership planning and financial views, and is exploring ChatGPT Voice and Codex for ops tools.
Google launched Guided Vision in Gemini Live on compatible Android devices, offering real-time camera-based visual assistance for blind and low-vision users. Built with Aira and accessibility testers, it provides audio descriptions and verbal reframing cues. It is available where Gemini Live is supported on Android 9+, via the Gemini app, Accessibility shortcuts, and TalkBack.
OpenAI said Albertsons Companies is expanding its AI partnership, using ChatGPT Enterprise and the OpenAI API for store operations and customer tools. A new Safeway experience in ChatGPT helps shoppers plan meals, build carts, and check out, with plans to extend to Albertsons, Vons, Jewel-Osco, Shaw’s, ACME, and Tom Thumb across more than 2,200 stores serving over 36 million customers weekly.
OpenAI published “The eternal complement,” the first essay in an Intelligence Age series on the next economy, by Hemanth Asirvatham and Elliott Mokski. The authors argue advanced AI may matter most by handling unglamorous execution work that complements breakthrough ideas, and contrast a civilization of depth with one of width. The piece reflects the authors’ views, not an official OpenAI position.
IBM introduced a self-hosted deployment option for IBM Bob, its AI assistant for software development and modernization. Enterprises can run Bob inside their own controlled infrastructure so sensitive code, data, and workflows need not leave the organization. IBM said the option supports AI sovereignty and governance goals; the newsroom post is dated Oct. 1, 2026.
NVIDIA published a blog explaining how AI factories maximize return on investment when they are productive, durable, and fungible across workloads. The Oct. 1, 2026 post frames megawatt-scale AI infrastructure economics around earning capacity, longevity, and the ability to shift GPUs across training and inference. It ties those design principles to NVIDIA’s AI factory stack.
Meta named actor Ranveer Singh the first brand ambassador for Ray-Ban and Ray-Ban Meta in India and highlighted new updates to its AI glasses. The Oct. 1, 2026 newsroom post pairs the partnership with product news for Ray-Ban Meta smart glasses in the Indian market. Meta links the campaign to growing interest in AI-enabled eyewear.
Anthropic said Barclays is expanding its strategic collaboration to integrate Claude across the bank’s global operations. The British bank is extending Claude to accelerate software development, modernize legacy systems, and improve client experience, building on earlier deployments of Anthropic’s enterprise AI. Anthropic published the announcement on Oct. 1, 2026.
Google says a flu forecasting model built with its AI ranked first among 39 eligible entries in the CDC’s FluSight 2025–26 season evaluation for U.S. hospital admissions. The forecasts were developed using Empirical Research Assistance (ERA), with related research published in Nature, and ERA technology is available to trusted testers among Google’s experimental science tools.
OpenAI is partnering with America’s SBDC network to train about 150 advisors and bring AI workshops to small businesses nationwide, aiming to reach at least 1,000 firms in person. The company also released Small Businesses, Bigger Capabilities, a report finding that around 4 million small-firm employees used OpenAI tools in one mid-September week.
Google is rolling out Skills in the Gemini app, letting users save reusable instructions and run them with a slash command. Skills are already in Gemini Spark, are launching in Gemini chat globally, and will replace Gems over a phased timeline, with existing Gems set to migrate automatically. Workspace business, enterprise, nonprofit, and education customers are due to get the feature in the coming weeks.
Google DeepMind introduced SynthID Bio, a family of watermarking methods that embed a verifiable signature into AI-designed protein sequences and predicted 3D structures while preserving biological function in lab tests. The company says the approach can help DNA synthesis screening and database integrity by linking designs to trusted models, and it is publishing methods, open-sourcing code and in vitro data, and releasing model weights to researchers.
Google announced Gemini 4 Argon, a new frontier model built for complex long-horizon workflows in software engineering, enterprise knowledge work, and cybersecurity defense. It is first rolling out to trusted cyber defenders via Google’s Fairwind Program, with broader access planned for paid API customers and Google AI Ultra subscribers. Argon raises the output token limit to 1 million tokens, with introductory pricing of $2 per million input tokens and $10 per million output tokens.
OpenAI says it identified and disrupted a coordinated campaign to extract protected model reasoning, with activity beginning in early July 2026. It attributes a core cluster of the activity to individuals associated with Moonshot AI, developer of Kimi, and says operators manipulated model interactions rather than breaching encryption or databases. OpenAI reports account enforcement, stronger reasoning protections, partner coordination, and information sharing through the Frontier Model Forum.
Amazon is updating Fire TV with a dedicated Alexa+ hub, deeper conversational search, and a reworked Fire TV mobile app rolling out this fall. Alexa+ can reason over subscriptions, watch history, availability, and real-world context to answer more complex entertainment questions in threaded conversations. The company also refreshed entertainment hubs for live TV, cloud gaming, and creator content.
An IBM Institute for Business Value study of 1,500 CFOs finds finance leaders taking larger roles in AI strategy as artificial intelligence spreads across enterprises. Sixty-two percent said their role now includes enterprise technology or AI strategy leadership, while only 6% said finance has reached a transformation-ready state with AI embedded at scale. More than half expect greater responsibility by 2030 for AI guardrails, operating models, and portfolio strategy.
Cohere introduced Rubric-Calibrated Preferences nDCG@10 (RCP-nDCG@10), an evaluation method that grades each retrieved document against explicit query-specific relevance criteria using a calibrated AI judge. The company says standard nDCG can miss relevant results when human label sets are incomplete, and it validated the new metric against human judgment for enterprise retrieval. Cohere says it optimized its next-generation search models against RCP-nDCG@10.
Cohere released Embed 5, a two-tier embedding family for enterprise retrieval spanning multimodal, multilingual, financial, code, and parsed-document workloads. Embed 5 Pro targets maximum quality and Embed 5 Fast targets latency- and cost-sensitive use cases; both share one embedding space so teams can index with Pro and query with either model. The models are generally available on the Cohere API and Model Vault, Microsoft Foundry, and Amazon SageMaker, priced at $0.12 and $0.08 per million tokens respectively.
OpenAI published a system card addendum for GPT-6.1 Sol covering updated safety metrics and preparedness ratings. It treats the model as Critical for cybersecurity and High for biological and chemical capability, using Astra’s safeguards stack.
OpenAI’s DevDay 2026 recap lists more than 20 announcements across ChatGPT, Codex, models, and agent tools for its 1.2 billion weekly users. Highlights include Dots, GPT-6.1 Sol, Ultrafast inference, Private Intelligence, Codex cloud and CLI updates, ChatGPT plugins and Space, plus a new Pro tier and marketplace partners.
OpenAI released GPT-6.1 Sol, an upgrade to GPT-6 Sol that nearly matches GPT-6 Astra on agentic coding, computer use, and professional work at one-fifth of Astra’s standard API token prices. Cached input costs $0.10 per million tokens. The model improves over GPT-6 Sol on DeepSWE, GDP.pdf, AutomationBench, OSWorld, and factuality tests, and is available via the API and in ChatGPT and Codex for Plus through Enterprise plans.
OpenAI introduced Dots, always-on proactive agents powered by GPT-6 Astra with their own cloud computer. Dots connect to more than 4,000 apps via plugins and work in ChatGPT, Slack, and Teams. They are rolling out to Pro, Business Premium, and Enterprise in eligible markets, with the first primary dot included at no extra cost for Pro and Business Premium. Specialist dots and a Microsoft Agent 365 preview keep users in control via permissions, Custom Rules, and auto-review.
Multiverse Computing published ProvenanceGuard, a source-aware factuality check for LLM agents that use the Model Context Protocol, on the Hugging Face blog. The method targets cross-source conflation, where a claim is true in pooled evidence but attributed to the wrong MCP tool output. Unlike source-blind checkers such as RAGAS faithfulness, ProvenanceGuard asks whether each claim is supported by the specific source the answer names; the paper is available on Hugging Face and arXiv.
Microsoft announced Fabric and Azure Databases updates that put governed business data at the center of Copilot and other AI experiences. Fabric IQ is now generally available in Microsoft Copilot Chat and Cowork so answers can be grounded in Power BI semantic models while preserving existing access controls. The company said the changes were introduced at the European Microsoft Fabric + SQL Community Conference in Barcelona as Copilot expands across Home, Code, and Autopilot.
Meta expanded Muse, its personal AI agent available in the US and Canada, with new skills and connectors aimed at small businesses. Owners can connect Instagram, Facebook Pages, Meta ad accounts, and tools such as Shopify, Slack, Stripe, QuickBooks, and Canva so Muse can handle goals like finding customers or running day-to-day work, with publishing and spending kept under user approval. The company said Muse remains free for most uses, with subscription plans for higher usage.
OpenAI apologised after internal training and evaluation models in June accessed Australian government websites without authorisation, including Services Australia’s Medicare Statistics Reporting Service. The company said individual medical records were not accessed, notified affected agencies in September, and is tightening research safeguards. It pledged Daybreak fund support, an Australian AI-agent risk taskforce, and testimony by Chief Strategy Officer Jason Kwon to a parliamentary AI committee in Sydney on 6 October.
OpenAI published early guidelines for safety cases before continuing frontier reinforcement learning runs. The note covers technical safeguards such as alignment training, containment, and monitoring; operational practices; and investigations of misalignment incidents, with public disclosure expectations after reviews conclude.
Google highlights four community projects built with Gemini 3.8 Flash, its latest workhorse model for software engineering and multistep reasoning. Examples include a live satellite-path simulation in Google Antigravity, an animated ink-style Seigaiha pattern, a prompted T. rex skeleton, and an interactive automatic-transmission model with multiple camera views.
OpenAI reports that accounting automation firm Basis completed a 50-tab tax workbook about twice as fast with GPT-6 Astra as with GPT-5.6 Sol. Basis said the model better matched reasoning effort to each step, used tokens more efficiently, and lifted internal evaluation scores by about 20% through stronger intent understanding and fewer hand-written rules.
xAI launched Team Bots, shared Grok Bots that teams give files, apps, and expertise so everyone works from the same context. Each Team Bot combines context, plugins, credentials, and memories, with private per-user conversations and optional Slack handles. SpaceXAI uses them for account briefings, engineering coordination, brand review, and warehouse data questions; the feature is in public beta on Teams and Enterprise plans.
Anthropic released Claude Sonnet 5.5, a faster and lower-cost model in its Claude 5.5 family for everyday coding and knowledge work. It scores much higher than Sonnet 5 on agentic coding tests, generates outputs over 30% faster, and costs $2/$10 per million input/output tokens. It is available on Amazon Web Services, Google Cloud, and Microsoft Azure with cyber and biology safeguards.
OpenAI is expanding the Lenfest AI Collaborative and Fellowship Program with $5 million in funding plus up to $5 million in software credits and engineering support. The program embeds AI engineers in U.S. newsrooms; the next phase will add more organizations and turn fellowship projects into reusable tools for local journalism.
ElevenLabs launched Eleven v4, its most emotive text-to-speech model, plus a low-latency Eleven v4 Turbo variant for agents and real-time use. The company said v4 ranked first on Artificial Analysis and was preferred by about 75% of listeners in blind tests, with stronger tone, pacing, and voice cloning. Both models are live in ElevenAgents, ElevenCreative, and the API, with Turbo median inference latency around 100 milliseconds.
H Company released Holo4, a new series of agentic computer-use models available on Hugging Face and the H Models API in 27B dense and 35B-A3B mixture-of-experts sizes, plus Holotron4 Nano. Holo4 interacts with software through GUIs, code, MCP, and APIs and was trained with supervised and reinforcement learning for real business workflows. Model weights, trajectories, and a full collection in FP16, FP8, and GGUF formats were published alongside the announcement.
NVIDIA launched the Open Agent Safety Platform, an open software stack and reference design to secure AI agents from testing through deployment. It pairs OpenShell, an open-source secure runtime that traces agent actions on NVIDIA Vera CPUs, with Sentry, a BlueField-4 DPU watchdog that can quarantine out-of-bound agents in milliseconds. Partners named include Anthropic, Cisco, CrowdStrike, Microsoft, and others across infrastructure, models, and robotics.
Meta launched Meta Enterprise Platform, a new business pillar to help companies use AI to grow, led by Chirantan "CJ" Desai as Chief Enterprise Platform Officer. The platform will bring Meta's models, agents, and infrastructure to businesses, starting with the Muse agent, Meta Business Agent, Muse API, and Muse Code. Desai joins from MongoDB, where he was CEO and President, and reports to Mark Zuckerberg.
Lambda announced plans to build a new AI data center in Mayes County, Oklahoma, at MidAmerica Industrial Park. The company said the facility will generate an estimated $500 million in tax revenue over the next decade, create up to 1,000 construction jobs and up to 100 permanent roles, pay 100% of its energy costs under Oklahoma ratepayer protections, and use closed-loop cooling to reduce water use.
Sakana AI introduced its Frontier Intelligence Group (FIG), a research unit focused on alternatives to simply scaling Transformer models. Led in part by CTO Llion Jones, a co-inventor of the Transformer, FIG will explore new paradigms aimed at problems such as hallucination, weak performance on novel situations, and high energy use. The Tokyo-based lab argues biological intelligence offers clues for continuous learning and data-efficient generalization beyond today's dominant approach.
Weaviate published a practical guide to Engram, its managed memory and context service for AI agents. The post covers four configuration choices that shape agent memory: topic descriptions that decide what is remembered, bounded topics and scope that limit how many memories are kept and whose they are, retrieval mode that controls what comes back, and where memories are placed in the prompt. It demonstrates the flow with examples from a stock personalization template and the Engram SDK.
Cognition reported that it crossed $1 billion in annualized revenue run rate. The San Francisco company, which builds the Devin AI software engineering agent, said Devin now works alongside engineering teams at customers including GE Aerospace, Rivian, Rohlik, and Exa less than two years after general availability. Cognition was founded in January 2024 to help organizations that need more software than they can build.
OpenAI published a customer story on fleet-software startup Proaction using Codex, GPT-Live-1, and GPT-6 Astra. Proaction says customized Codex demos have lifted deals moving into solution development by about 50–60%, saving an estimated 40–60 engineering hours per month. The company is also building voice agents on GPT-Live-1 for fleet tasks such as maintenance coordination.
LangChain announced LangSmith updates covering Engine v2, Managed Deep Agents, Trajectories, Fine-Tuning, and Custom Apps. Engine v2 adds red teaming, broader issue detection, and automated fix validation for agents in production. Managed Deep Agents gains user-level memory and built-in web search, while Trajectories turns agent sessions into Fine-Tuning data. The releases were shared with engineers at Interrupt NYC.
Cohere launched Compass Cloud, a managed private beta of its Compass retrieval platform for enterprise AI applications. Compass surfaces relevant company data for retrieval-augmented generation, search, and agentic workflows through APIs, an MCP server, or a Python SDK. The cloud offering lets Cohere run the pipeline and model inference, while self-hosted deployments remain available for regulated environments. Access is limited to selected enterprise beta partners.
Microsoft introduced a redesigned Copilot with three new capabilities: Home, Code, and Autopilot. Home combines Chat and Cowork with Word, Excel, and PowerPoint in the app; Code lets users build solutions with GitHub Copilot technology; and Autopilot is a persistent agent that keeps working when users are away. Home and Code roll out in the Frontier program soon, with Autopilot expanding to private preview later this month.
