Model launches led 18 AI releases on 7 October. OpenAI began rolling out GPT-6 to all ChatGPT users and published an updated system card, while Anthropic released Claude Haiku 5.5 for high-volume, cost-sensitive work. Google opened its SynthID Detector to everyone globally, and Microsoft and NVIDIA outlined on-device AI agents and Copilot coding for new Windows PCs.
Wednesday 7 October 2026 (America/Los_Angeles source-local date) · 18 releases · MediaRelease.co editorial digest
LangChain has revamped Agent Skills support in Deep Agents as enterprise skill registries grow to thousands of skills. Tools bound to a skill now load only when the agent reads that skill, keeping context small. Applications can also preload explicitly requested skills before the first model call, and long-running agents can pick up added, edited or deleted skills without starting a new thread.
Providence, a US health system with 119,000 caregivers, worked with IBM Consulting to simplify hiring and internal transfers using AI and automation without replacing existing systems. Hiring managers cut administrative effort by 90%, internal caregivers now start new roles an average of 12 days sooner, and transfer costs fell 60%. Nearly half of Providence's annual hiring comes from internal transfers.
Lambda has summarised ReLearn, a CVPR 2026 workshop it co-organised and sponsored that asked whether AI can still learn from human intelligence. The workshop brought together researchers from computer vision, cognitive science and embodied AI, and Lambda provided compute credits for every accepted paper. Invited speakers included Alexei Efros, Manling Li, Dima Damen, Alan Yuille, Saining Xie and William T. Freeman.
SambaNova has made prompt caching available for MiniMax M3 on SambaCloud, with no code changes required. Requests sharing a stable prefix of at least 4,096 tokens, up to 192,000 tokens, can be served from cache, and cached tokens cost $0.06 per million instead of $0.60. SambaNova reports that cache hits cut time to first token by 35% to 88% across 8,000 to 192,000 tokens of context.
Salesforce has outlined recommendations for building trust as enterprises move from standalone large language models to interoperable multi-agent systems that work across company boundaries. The guidance comes from workshops with engineers, researchers, security specialists and ethics professionals from 21 organisations and is summarised in a new white paper. Salesforce argues that blurred security perimeters and latent malicious prompts make Zero Trust, with continuous verification of every interaction, essential.
Microsoft says GitHub Copilot will automatically decide whether a coding task should run on local on-device models or cloud-scale models, with the capability due by the end of the month. It will enable local coding on NVIDIA RTX Spark Windows PCs such as Surface Laptop Ultra. Microsoft also highlighted Microsoft Execution Containers for securing agentic coding sessions and Project HydraFusion, which selects models by balancing performance, cost and latency.
NVIDIA founder Jensen Huang and Microsoft CEO Satya Nadella discussed bringing AI agents to Windows PCs during a fireside chat at Microsoft's Windows AI and Surface event in San Francisco. NVIDIA said the two companies are co-engineering hardware and software so agents can run on Windows PCs, including systems with RTX Spark. Huang traced NVIDIA's decades-long history with Windows and described the PC becoming a personal assistant in the agent era.
Meta has released a conversation between developer and creator Tom Shaw and Meta Head of Infrastructure Santosh Janardhan about how the company builds infrastructure for AI. Janardhan discusses why Meta is more than a software company and why data centres are essential for AI. Topics include how Meta powers its infrastructure, what a gigawatt of energy is, how it chooses chips and the benefits of building its own data centres.
OpenAI published a system card update for the October 2026 versions of GPT-6 Sol and GPT-6 Luna, released as GPT-6 rolls out to all ChatGPT free and paid plans in place of GPT-5.6 Sol and Luna. OpenAI rates both models High capability in cybersecurity and in biological and chemical domains, but below its High threshold for AI self-improvement. It reports stronger jailbreak resistance than GPT-5.6, with some safety regressions it judged generally low severity.
OpenAI published a customer story on how Radisson Hotel Group worked with Accenture Song to build a Radisson Hotels plugin for ChatGPT that lets travelers compare hotels on a map and continue to Radisson’s site to book. Built in six weeks, the plugin delivered visit-to-booking conversion about 1.5 times Radisson’s organic search in July–August 2026. Radisson is also running sponsored ads in ChatGPT, with more than half of recorded checkout and booking events attributed through view-through measurement.
OpenAI shared early usage data for ChatGPT for Teens, its default experience for accounts identified as under 18, and said it will soon add College Planner to help US students in grades 10–12 track application requirements, deadlines and financial-aid steps. The company is also supporting the nonprofit College Advising Corps and a Boston Children’s Hospital Digital Wellness Lab student advisory council over three years. Flashcards, easier quiz creation and multi-photo note capture on iOS are also being added.
Anthropic released Claude Haiku 5.5, a small model aimed at high-volume, cost-sensitive work such as summaries, classification, subagent tasks and browser use. Anthropic says it costs about 75% less to run than Haiku 4.5 on average and is its first Haiku-class model with an adjustable effort setting. It is available on the Claude Platform, Amazon Web Services, Google Cloud and Microsoft Azure. Anthropic also halved Sonnet 5.5 cache-read prices and is adding monthly API credits for Max and Team subscribers.
OpenAI is rolling out a new GPT-6 model in ChatGPT, extending its next-generation models beyond paid customers to the more than 1.2 billion people who use the service each week. The update adds Intelligent UI, which lets answers include interactive charts, forms, maps and small tools, and lets ChatGPT begin answering while it is still thinking. Plus, Pro, Business and Enterprise users get GPT-6 Sol first, with Free and Go users moving to GPT-6 Luna from 8 October.
CoreWeave has introduced CoreWeave Direct-to-Expert, a support model that puts customers in a shared channel with engineers cross-trained across storage, networking and scheduling, with a human responding in minutes. Engagements typically begin in CoreWeave ARENA, a lab for testing real AI workloads on production infrastructure, and continue through onboarding, migration and production. The company also offers Physical AI Field Engineering and says 9 of the 10 leading foundation model providers rely on CoreWeave.
IBM has announced IBM Ready for SAP Solutions to help midsize and fast-growing organizations accelerate SAP Cloud ERP modernization, unify enterprise data and establish an AI-ready digital core. Volumetric Building Companies cut operational cycle times by 50% after a three-month transformation, while Second Nature Brands integrated its acquired Voortman Bakery business. The offering draws on IBM Consulting Advantage, IBM's AI-powered delivery platform, and is currently available in limited countries and industries.
Meta has outlined additional measures against child sexual exploitation on its platforms, including new large language model detection for ads that covertly signpost illegal material, AI-driven sweeps of ad content, a red-teaming AI agent and stronger recidivism detection. Meta said it actioned 33.2 million pieces of child sexual exploitation content on Facebook and Instagram globally between January and June 2026, including 5.3 million in India, most of it found proactively.
Google is expanding access to its SynthID Detector, making the tool available to everyone globally in English from October 7, 2026. Anyone can check whether an image, video or audio file was made with AI from Google or partners including OpenAI, NVIDIA and Kakao, with Apple to follow. Google says SynthID has watermarked more than 180 billion images and videos and 240,000 years of audio since its 2023 launch.
Google has launched Playground, an experimental browser-based platform that lets people create, play and share custom games by describing them in text prompts, with no coding required. Players can publish games to a community gallery that applies safety screenings, and compete through leaderboards and multiplayer in select genres. Playground is available to US users aged 18 and over at playground.google, with creation access tiered by Google One membership and a Unity Spark integration planned.
We are gauging interest in a daily or weekly newsletter for each area. Tell us which areas you would read.
