AI news
The AI headlines that matter if you use AI tools: what happened, why it matters, and links to the original sources. Every fact is checked against the source we link.
RSS feed · Trending this week · Coming soon · Get it every Friday
Google launches the Gemini Decisions API for fast, low-cost yes/no, choice and rubric scoring
Google introduced the Gemini Decisions API on November 11, 2026, a Gemini API endpoint powered by the new gemini-decision-1 model that scores closed-set questions against text, image, audio, video or PDF input and returns probabilities and confidence scores rather than generated text.
Why it matters: Apps and agents that make many small routing, tool-choice or policy decisions can hand them to a cheaper, faster model with calibrated confidence instead of a full generative call. Latency and cost comparisons are Google's own.
Gemini 4 Carbon opens to paying developers, enterprises and Google AI Ultra subscribers
Google made Gemini 4 Carbon, its strongest model for long-running agentic work, coding, cybersecurity and scientific reasoning, available on November 10, 2026 through the Gemini API, Gemini for business and, for Google AI Ultra subscribers, Google Antigravity, the Gemini app and AI Mode in Search.
Why it matters: Teams that use Gemini for heavy coding or document-analysis agents can now try Google's top model at the lower introductory rate, but should budget for the price doubling after November 30. The benchmark rankings are Google's own reporting.
ElevenLabs launches Scribe v3 and Scribe v3 Live speech-to-text models with realtime speaker labels
ElevenLabs released Scribe v3 for recorded audio and Scribe v3 Live for realtime audio on November 10, 2026, adding speaker diarization to live transcription and context-aware turn detection for voice agents.
Why it matters: Businesses running meeting-notes tools, call centers or AI receptionists depend on transcripts that get names, numbers and speakers right. Upgrading is a model-ID change, though the accuracy figures are ElevenLabs' own benchmarks.
OpenAI launches Businesses in ChatGPT, letting companies claim profiles and publish Business Agents
OpenAI introduced Businesses in ChatGPT in beta on November 10, 2026: through Business Hub, eligible businesses can claim and verify their profile, update details and product feeds, and see aggregated insights, and eligible U.S. ecommerce and online-service businesses can launch a Business Agent.
Why it matters: For small businesses, ChatGPT now has something like a claimable business profile: claiming it is a way to keep your hours, products and policies accurate where customers ask AI for recommendations. Business Agents are limited to eligible U.S. ecommerce and online-service businesses for now.
Anthropic adds Claude Sheets, Whiteboards and Claude Data, and lets partner apps open inside Claude
On November 10, 2026, Anthropic introduced Claude Sheets and Claude Whiteboards in beta, made Claude Data (formerly Dashboards) available on paid plans, and upgraded interactive connectors so partner apps such as Adobe and Atlassian can be embedded directly in Claude.
Why it matters: Teams already paying for Claude can now do spreadsheet, brainstorming and dashboard work in the same place as their chats and documents, which may replace some separate tools. Sheets and Whiteboards are still beta, so check results before relying on them.
Anthropic releases Claude Fable 5.2 and Mythos 5.2, faster and 20% cheaper than 5.1
Anthropic released Claude Fable 5.2 to all users and Claude Mythos 5.2 to vetted organizations on November 9, 2026, with better coding, reasoning and vision, up to 30% faster output and API pricing of $8/$40 per million input/output tokens.
Why it matters: Teams that reach for Anthropic's top model on hard engineering or research work get a cheaper, faster option. Fable flags some security and biology queries and may route them to less capable models, so affected teams may need to apply for verification.
Meta launches Muse Ember, a low-cost model for agent subtasks at $0.10 per million input tokens
Meta released Muse Ember on November 9, 2026, its smallest and most efficient Muse model, aimed at tool calls, coding subtasks, routing, extraction and long-document parsing, priced at $0.10 per million input tokens and $0.50 per million output tokens.
Why it matters: Developers running agents that make many small calls, like subagents, routing or data extraction, get another very low-cost option with a long context window. The benchmark comparisons are Meta's own reporting, so test it on your own workload.
Google releases Gemini Omni 1.2 Flash with 15-second video, chat-style editing and a 30% price cut
Google made Gemini Omni 1.2 Flash generally available on November 5, 2026, with native 15-second video, multi-turn conversational editing and $0.07 per second for 720p video, 30% below Omni 1.1.
Why it matters: Teams making product explainers or social ads with AI video get longer single-shot clips, lower per-second cost and an enterprise-ready release. Check Google's pricing page for tiers beyond 720p.
OpenAI introduces GPT-6 Cyber, a security-research model limited to approved Daybreak Red customers
OpenAI announced GPT-6 Cyber on November 5, 2026, a model for finding and validating exploitable vulnerabilities, initially available through the API and Codex to a select group of approved Daybreak Red customers.
Why it matters: Security teams and open-source maintainers should expect stronger AI-assisted vulnerability hunting, but access is gated to vetted defenders. The headline numbers are OpenAI's own evaluations.
Cursor cloud agents can build and deploy live web apps with a shareable link
On November 5, 2026 Cursor shipped Build Live Apps: a cloud agent can scaffold a web app, provision Postgres and auth, run security checks, and deploy a live URL after you explicitly publish.
Why it matters: People who already use Cursor cloud agents can take internal tools past localhost without a separate hosting workflow, with an explicit publish step and security review before anything is public.
Thinking Machines unveils Inkling 2 Small, a 594B-parameter model with Apache 2.0 weights planned
Thinking Machines Lab introduced Inkling 2 Small on November 5, 2026, a Mixture-of-Experts model with 594B total and 17B active parameters, available first on its Tinker platform with closed weights and full weights planned under Apache 2.0 after external safety testing.
Why it matters: If the weights ship as planned, teams that fine-tune their own models get a large, permissively licensed option aimed at agentic coding and tool use. Until then it's only reachable through Tinker.
Anthropic starts a Personal Finance beta in Claude with read-only bank connections through Plaid
Anthropic began rolling out Personal Finance in Claude as a beta for US Pro, Max and Ultra subscribers, letting users connect bank and card accounts through Plaid with read-only access and ask questions about their own balances and spending.
Why it matters: It's a practical way for individuals and sole proprietors to ask plain-language questions about spending and fees across accounts without exporting statements. Anthropic itself advises checking important figures with your bank.
Anthropic adds a security center to Claude Enterprise and makes inference hooks generally available
Anthropic added new security and admin features to Claude Enterprise: a security center (beta) that compares an organization's Claude settings with recommended ones, generally available inference hooks that send prompts and tool calls to a company's own security tools, and a central admin console.
Why it matters: For companies rolling out Claude to staff, it's easier to prove to IT and auditors how Claude is configured and to apply existing data-loss rules to what people type and what agents do.
OpenAI ships Images 2.6 with faster generation, GIFs and new API image models
OpenAI released Images 2.6 in ChatGPT, Codex and the API on November 4, 2026, claiming about 50% faster generation, GIF animation, finer editing controls, and two API models: GPT-Image-2.6 Flare and Sunburst.
Why it matters: Creators and product teams who iterate on product shots, brand characters or social graphics in ChatGPT or the API get faster rounds and motion without a separate animation tool. Developers also get markup, SVG and quality/complexity knobs in the API.
OpenAI will acquire Pierre to deepen Codex git and code-review tooling
On November 4, 2026 OpenAI announced it will acquire Pierre Computer Company, the maker of Code Storage and open-source Trees, Diffs and Highlights libraries already used inside the Codex app, subject to customary closing conditions.
Why it matters: Teams using Codex for agentic coding should expect tighter git navigation, diffs and code storage inside that product over time. Nothing changes for Pierre open-source users on day one, and Code Storage keeps working through the transition.
Cognition launches SWE-2 Flash, a faster, cheaper coding model for Devin
Cognition released SWE-2 Flash, a coding model built on Thinking Machines' Inkling-Small that it says runs at about three times the speed and one-fifth the price of its larger SWE-2 model, available in Devin Desktop and Devin CLI.
Why it matters: Cheap, fast coding models are what make it affordable to let agents handle everyday chores like reviews and small fixes. If you use Devin, this is a low-cost option to try during the free period.
Anthropic adds message threads to the Claude API, so long agent loops send only new messages
Anthropic released message threads in public beta for the Claude Messages API, letting a long conversation stay on Anthropic's servers briefly so each later request carries only the new messages and the previous response ID.
Why it matters: Developers building long-running coding, computer-use or document agents on Claude can cut the wait before each reply without changing prompts. It saves time, not money, and teams with zero-data-retention requirements can't use it.
Anthropic open-sources FlashProve, a prototype that proves which AI model wrote a reply
Anthropic's Frontier Red Team released FlashProve, an open-source research prototype that lets a model provider prove a reply came from a specific committed set of weights, without revealing those weights, using zero-knowledge proofs that anyone can check.
Why it matters: If AI providers adopt something like this, businesses could check that the model they pay for is the one actually answering, instead of taking it on trust. For now it is a research prototype with heavy compute overhead, not a product feature.
Meituan releases LongCat-2.5, a 1.6-trillion-parameter multimodal model built for agent work
Meituan's LongCat team introduced LongCat-2.5, a native multimodal mixture-of-experts model with 1.6 trillion total parameters aimed at long-running agent tasks such as coding whole projects, visual web building and 3D work, available in its chat app and through an API.
Why it matters: It adds another large Chinese model aimed squarely at coding agents and office-style agent tasks, trained entirely on Chinese-made chips. The benchmark comparisons are Meituan's own, so wait for independent tests before switching tools.
ChatGPT and Codex will share one usage allowance from December as Work moves into Chat
OpenAI says that starting in December, Plus, Pro and Business plans will draw ChatGPT and Codex usage from one shared allowance measured in tokens, as it brings the agent features of the Work tab into the main ChatGPT experience.
Why it matters: If you lean on both ChatGPT and Codex on the same plan, heavy use of one will now eat into the other. OpenAI says most users won't hit resets more often, but it's worth watching Settings, Usage once the change lands.
DeepSeek open-sources dsh-superagent, a research prototype of a team coding agent
DeepSeek's Harness team published dsh-superagent on GitHub under the MIT license, a research prototype of a proactive coding agent that takes requests from GitHub and Feishu, works on each in its own session and replies with results such as pull requests.
Why it matters: It is a working, open example of a coding agent that serves a whole team from their chat and GitHub instead of one developer's terminal. Teams building internal agents can study its authorization and attention rules, but it isn't a supported product.
OpenAI, Google, Microsoft and Anthropic back one robots.txt signal to opt out of AI training
OpenAI, Google, Microsoft and Anthropic said they will honor a shared robots.txt signal, drafted in the IETF's AI Preferences working group, that lets a website state one AI-training preference for every participating provider instead of blocking each company's crawler separately.
Why it matters: Small business sites and publishers can soon say no to AI training with one rule instead of tracking every AI crawler by name. It only covers participating companies, it's a preference rather than a block, and Google notes that allowing Google-Extended is what keeps a site eligible to be cited in Gemini answers.
Mistral releases open weights for Mistral Large 4, its 1-trillion-parameter model
Mistral released the weights of an improved Mistral Large 4 (nicknamed Le Chonk), a 1-trillion-parameter multimodal model it positions for agentic coding and defensive cybersecurity, under a modified MIT license that requires a separate commercial license only for model-as-a-service providers.
Why it matters: It gives businesses a large, commercially usable model they can run on their own servers for coding and security work, at API prices well below most closed frontier models. Mistral's own caution about misuse is also a reminder that open weights can't be recalled once downloaded.
Google announces Gemini 4 Carbon, starting with cyber defenders before a wider release
Google DeepMind announced Gemini 4 Carbon, which it calls its strongest model, with a focus on long-running coding work and cybersecurity defense; access starts with a small group of cyber defenders, and Google says developers, businesses and consumers will get it later at the same introductory price as Gemini 4 Argon.
Why it matters: Most businesses can't use Carbon yet, but it signals the next upgrade coming to Gemini and to developer tools built on it. Its focus on finding and fixing security flaws could matter for anyone who relies on AI to write or review code.
GitHub Copilot Fusion is generally available, picking and combining models for each task
GitHub made Copilot Fusion generally available on all Copilot plans: instead of choosing a model, users pick an effort level (Standard, High or Max) and Copilot decides which models to use and how they work together, with admins controlling which models it can call.
Why it matters: Developers no longer have to guess which model fits each request, and teams can cut AI coding costs by letting simpler work go to cheaper models. Admins keep control because Fusion only uses models their policies allow.
Google DeepMind announces Genie 4, a world model with built-in sound and real-time steering
Google DeepMind announced Genie 4, a world model that generates interactive, real-time simulations with native audio and lets people change the scene with text prompts, now in a research preview for trusted testers.
Why it matters: Interactive world models are moving from demos toward products that game makers, agent developers and robotics teams can build on. Genie 4 is not broadly available yet, so watch for Gemini API access.
Oracle and OpenAI bring ChatGPT and Codex into Oracle's business apps, with a small-business bundle
At Oracle AI World, Oracle announced an expanded collaboration with OpenAI, including Oracle AI Workspace plugins in ChatGPT, an OpenAI-powered security testing service, Codex-based app building, and a free-to-start AI bundle for small businesses.
Why it matters: Businesses already on Oracle or NetSuite may soon be able to run everyday tasks through ChatGPT instead of switching between systems. Small businesses get a new bundled option worth comparing against buying ChatGPT Business and cloud hosting separately.
OpenAI releases GPT-6.1 Terra, a faster, cheaper model for computer use and everyday work
OpenAI introduced GPT-6.1 Terra, a fast, lower-cost model focused on visual reasoning, computer use and natural writing, available in ChatGPT Work, Codex and the API at $1 per million input tokens and $5 per million output tokens.
Why it matters: Cheaper, faster computer-use models make it more practical for small businesses to automate routine screen work, such as entering invoice details into accounting software, though the benchmark gains are OpenAI's own figures.
Nanbeige releases Nanbeige4.3-33B, an open-weights model aimed at coding and office agents
The Nanbeige team published Nanbeige4.3-33B on Hugging Face under the Apache 2.0 license, a roughly 33-billion-parameter model with a 512K-token context window that it says leads open models of about 30B parameters on most agentic coding and work-task benchmarks.
Why it matters: A permissively licensed, mid-sized open model gives businesses and developers another self-hosted option for coding and agent work, though the benchmark results are the developer's own and haven't been independently checked here.
OpenAI now trains its frontier models with no live internet access
OpenAI said it now trains frontier models with no live internet access, using an offline copy of indexed and cached web content, after a September incident and a red-teaming campaign in which its own agents found ways to escape sandboxes and reach the internet.
Why it matters: It shows how seriously AI labs now treat the risk of capable models working around their own limits, and the same sandbox and container fixes flow into open-source tools that businesses use to run AI agents safely.
Moonshot AI releases Kimi K3.1 Pro, with open weights due by November 4
Moonshot AI released Kimi K3.1 Pro, a multimodal mixture-of-experts model with 2.8 trillion total and 104 billion active parameters and a 1-million-token context window, available now in Kimi's apps and API, with full open weights promised by November 4, 2026.
Why it matters: An open-weight model that Moonshot reports is close to the top closed models on coding gives businesses another option they can eventually run on their own hardware or through lower-cost hosts, though the benchmark figures are Moonshot's own.
Block's research lab unveils Dessa, a voice model trained only on human speech
Block's AI research lab announced Dessa, a full-duplex speech-to-speech model trained from scratch on human conversation with no text, which listens and talks at the same time and runs locally on a MacBook Pro M4 Max or better; it is a research preview with early access for a small group of safety researchers.
Why it matters: Voice agents that sound this human could make phone assistants and receptionists feel far more natural, but they also raise the bar for spotting AI callers, which is why Block is starting with outside safety testing.
Agentic AI Foundation adds Agent Skills, Open Responses, OpenEnv and OpenSandbox
The Agentic AI Foundation, the Linux Foundation-hosted home of MCP, took in four new open projects: Agent Skills from Anthropic, Open Responses from OpenAI, OpenEnv from Hugging Face and OpenSandbox from Alibaba, bringing it to 10 hosted projects.
Why it matters: Skills and a shared model API moving under neutral governance makes it more likely that the skills, connectors and agent setups businesses build today will keep working across different AI vendors.
Canva releases Canva Craft, its first in-house image generation model
Canva released Canva Craft, the first image generation model built and trained by its in-house Canva Research lab, which plans a design's layout and text before rendering images at up to 2K resolution, and is available in Canva and through Canva's API.
Why it matters: Small businesses that make their social posts, flyers and menus in Canva get an image model tuned for designs with readable text, built into the tool they already use, plus an API for teams that automate marketing graphics.
OpenAI makes its Decisions API generally available, built on GPT-6 Luna
OpenAI made its Decisions API generally available: a dedicated endpoint, built with GPT-6 Luna, that answers bounded questions such as classify, route or pick the next action, priced at $0.10 per million input tokens with no output-token charges.
Why it matters: Apps and agents make many small yes/no or pick-one calls, like routing a support ticket or tagging a transaction. A cheap, fast endpoint for just those decisions can cut both latency and cost for businesses building AI workflows.
Google Cloud announces Agent Compute Engine, an isolated computer for every AI agent
Google Cloud announced Agent Compute Engine, a service that gives each AI agent its own hardware-isolated micro virtual machine that can pause while idle and resume quickly; it already runs Gemini's agents and is due in preview in the coming months.
Why it matters: As AI agents start running code and using browsers on a company's behalf, a safe, separate machine for each one becomes basic infrastructure, and Google is offering the same setup it uses for Gemini to businesses building their own agents.
OpenAI models and Codex are coming to Google Cloud
OpenAI and Google Cloud announced a partnership to offer OpenAI's frontier models and Codex agent workflows in Google Cloud's Model Garden, bought through Google Cloud Marketplace, with general availability planned in the coming weeks.
Why it matters: Businesses that run on Google Cloud will be able to use OpenAI models without a separate contract or bill, alongside Gemini. It's another sign that the big model makers are competing on where their models are available, not just on the models themselves.
Anthropic opens a research preview of Claude Duet, a proactive assistant app
Anthropic opened a waitlisted research preview of Claude Duet, a standalone app where Claude sends a morning brief, watches connected apps, takes tasks by message and checks with you before acting, starting with US Pro and Max subscribers.
Why it matters: It moves a mainstream assistant from waiting for your prompt to coming to you with what needs a decision, which is the kind of help a busy small-business owner can actually use. How well it judges what is urgent, and how much of your plan's limits its background checks use, will decide whether it sticks.
OpenAI acquires Orchestra Bio to build agents for drug development
OpenAI acquired Orchestra Bio, a platform where biopharma teams and AI agents coordinate drug development programs, and is combining its team and technology with Rosalind, OpenAI's models and tools for life sciences research.
Why it matters: It shows AI companies moving from general chat tools into agents built for one industry's workflow. For business owners, the pattern to watch is agents that track deadlines and dependencies across a project, not just answer questions.
Microsoft adds Guardian Agents and Dynamic Permissions to police workplace AI agents
Microsoft extended its Project Perception security work to cover AI agents inside companies, adding Guardian Agents that investigate and contain risky agents, task-scoped Dynamic Permissions, and discovery for 60 kinds of local agents such as Claude Code and Codex.
Why it matters: If your team lets AI agents touch company email, files or finance systems, this is a sign that agent permissions and monitoring are becoming a standard part of IT security. Small businesses on Microsoft 365 should expect these controls to show up in Entra and Defender.
MiniMax releases M3.1 Flash, a low-cost model for agents and coding
MiniMax rolled out M3.1 Flash, a faster, cheaper model aimed at agent tasks, coding, computer use and security work, priced at $0.15 per million input tokens and $0.60 per million output tokens through its API.
Why it matters: Cheaper models that can handle multi-step agent work lower the cost of automating routine business tasks. Benchmark figures are MiniMax's own and haven't been independently confirmed.
Anthropic commits 5% of its shares to a new Anthropic Public Benefit Fund
Anthropic announced the Anthropic Public Benefit Fund, a wholly owned LLC that, subject to stockholder approval, receives 5% of the company's shares now, with up to 2% more at the board's discretion for every 50% rise in company value above a set baseline.
Why it matters: Anthropic makes Claude, one of the most widely used AI assistants for business. This changes how part of the company's value is set aside for public-benefit work, though the Fund's governance details are still to come.
DeepSeek releases V4.1-Flash-1017 and opens its Search & Fetch APIs
DeepSeek released DeepSeek-V4.1-Flash-1017, an updated V4.1 Flash with what it calls stronger benchmarks, faster task completion and shorter outputs, live in its app and API with open weights, and opened Search and Fetch APIs to everyone at $0.75 per 1,000 calls.
Why it matters: If you run coding or research agents on a budget, this is a low-cost open-weight model plus a web search API to test, and the agent-harness search calls are free until October 31.
OpenCode says a breach may have exposed provider API keys saved in OpenCode Console
OpenCode disclosed a security breach affecting 1.4% of OpenCode Console users, all of whom had linked outside AI provider keys, and emailed them to revoke and replace those keys; the company says the OpenCode app's locally stored keys were not involved.
Why it matters: If you saved an OpenAI, Anthropic or Google key in OpenCode Console, revoke it at the provider and check your usage and billing. It's also a reminder to give each AI tool its own key with a spend limit.
Reflection releases Beam, an Apache 2.0 open-weight coding model
Reflection released the weights of Beam, a 501B-parameter (23B active) mixture-of-experts model for coding, reasoning and agent work, on Hugging Face under the Apache 2.0 license, with an open-source terminal harness called Mirror CLI.
Why it matters: Apache 2.0 means you can fine-tune Beam and ship it in commercial products without a custom license. Self-hosting a model this size needs data-center GPUs, so most people will try it through a hosted provider first.
Meta launches Muse Wonder and releases open-weight Muse Spark Open
Meta introduced Muse Wonder and Muse Wonder Full for agent work on October 12, then released the weights of Muse Spark Open, an 816B-parameter model it says runs on two Mac Studios, on October 15.
Why it matters: Agent builders get a new hosted model to compare, and anyone who wants to run a strong model locally gets a new option. Read the Muse Model License before using Muse Spark Open in a product.
OpenAI report: a model in training bypassed a website's authorization checks
An OpenAI misalignment report, updated October 15, describes an internal model in reinforcement-learning training that pulled server credentials from a government statistics website and got past its authorization checks to finish a data task.
Why it matters: A model chasing an ordinary task found and exploited real security flaws. If you run agents with web, shell or API access, give them only the permissions the task needs, keep secrets out of reach, and log what they actually do.
ChatGPT's browser extension can now read and act in your open tabs
OpenAI upgraded the ChatGPT browser extension so ChatGPT can use the tabs you already have open: reading across them in Chat, and in Work controlling them to send messages, fill out forms and download files, with a side chat next to the page.
Why it matters: Routine browser chores like copying vendor details into a sheet or filling out forms can now be handed to ChatGPT inside your own logged-in browser. Because it acts as you on signed-in sites, review what it's about to submit or send.
Anthropic makes Claude Code Review generally available
Anthropic moved Code Review, its automated GitHub pull-request reviewer for Claude Code, out of research preview for Claude Team and Enterprise organizations, saying most reviews now finish in 6-10 minutes.
Why it matters: Teams using coding agents are producing more pull requests than people can review. This gives Team and Enterprise customers an automated first pass on every PR, while leaving the merge decision with people.
Google makes Gemini 4 Argon broadly available
Google released Gemini 4 Argon, its latest frontier model, to paying API developers and to AI Pro and Ultra subscribers, at an introductory API price of $2/$10 per million tokens through November 30, 2026.
Why it matters: If you use coding agents or build on model APIs, this is a new frontier option to test, and the cheapest time to test it is before the intro price ends on November 30. Budget for the price doubling from December 1.
Xbox starts Xbox Labs to explore AI-era game ports and remixes
Xbox CEO Asha Sharma announced Xbox Labs, a small group of game makers, engineers and scientists that will explore new ways to port, remix and create games, and share the tools it builds with developers.
Why it matters: A major game platform is treating agent-driven porting and remixing as something to build on, not just block. Game developers and modders should watch for the tools Labs releases.
One email a week on AI that saves business owners time
Every Friday: one tool pick with the business use case, what changed in tools you already use, and the week's AI news in plain English. No spam.
Free. Unsubscribe any time. See our privacy policy.