{"title":"Weekly Edition · 9–15 November 2026","description":"The Weekly Edition — Brew Bytes' curated roundup of the week in AI.","url":"https://brewbytes.io/news/week/2026-11-09","feedUrl":"https://brewbytes.io/api/functions/weeklyFeed?week=2026-11-09","weekStarting":"2026-11-09","weekEnding":"2026-11-15","weekRange":"9–15 November 2026","updatedAt":"2026-11-20T15:41:06.603Z","edition":null,"itemCount":7,"items":[{"id":"2026-11-15-gemini-4-hydrogen-replaces-3-8-flash","url":"https://brewbytes.io/news/item/2026-11-15-gemini-4-hydrogen-replaces-3-8-flash","headline":"Google released Gemini 4 Hydrogen at the price Gemini 3.8 Flash moves to in January, and recommends it as the upgrade","topic":"Models","date":"2026-11-15","whatItMeans":"On 6th Nov'26 we covered Google routing gemini-2.5-pro calls to 3.8 Flash from 1st Dec and doubling 3.8 Flash's price a month later. Hydrogen changes what to test in that window, because from January it costs the same as 3.8 Flash and Google now recommends it instead, so accepting the reroute means settling on a model Google has already moved past, at the same price.\n\nIf you call Gemini through the API, run your evaluation set against gemini-4-hydrogen as well as 3.8 Flash before 1st Dec, and set the model name yourself rather than leave the reroute to choose for you. Check the bill as well as the answers, because Hydrogen defaults to a higher thinking level than 3.8 Flash and bills thinking tokens as output. Code that sets a temperature or uses explicit caching needs changing before you switch.","body":"On 12th Nov'26 Google released Gemini 4 Hydrogen, the fast, cheaper model in its Gemini 4 family, aimed at agent work and everyday tasks. It is live as a stable model (gemini-4-hydrogen) on the Gemini API and AI Studio, on free and paid tiers, and in Search and the Gemini app. It costs $1.50 per million input tokens and $7.50 per million output tokens, which is what Gemini 3.8 Flash will cost from 1st Jan'27, and Google's migration guide calls it the recommended upgrade from 3.8 Flash. Google says it finishes tasks 2.5 times faster than 3.8 Flash. Its default thinking level is high where 3.8 Flash's is medium, and the guide says to remove custom temperature settings and explicit context caching, which Hydrogen does not support. Google has not announced a retirement date for 3.8 Flash.","theLine":"Before 1st Dec, test the Gemini model Google now recommends, not the one your old calls get rerouted to.","teaser":"","source":{"name":"Google AI Studio","url":"https://aistudio.google.com/docs/whats-new-gemini-4-hydrogen"},"isLead":true,"isSharable":false,"keyFigure":"","figureLabel":""},{"id":"2026-11-14-conductor-reverses-meta-shutdown","url":"https://brewbytes.io/news/item/2026-11-14-conductor-reverses-meta-shutdown","headline":"Conductor said it was joining Meta and closing its cloud service, then reversed course nine days later","topic":"Tools","date":"2026-11-14","whatItMeans":"Most teams pick an agent tool on how well it runs the agents, and few check what happens to their workspaces if the company behind it changes hands. Conductor gave its users nine days' notice of a cloud shutdown and then took it back, so the part worth looking at is the exit it published in between, because that is what you would have relied on.\n\nIf you run coding agents through a wrapper like this, keep the work itself in your own git repository and treat the tool's chats and workspace history as disposable. Then find the export option in each tool you use and run it once now, so you know what comes out and in what form before you need it.","body":"On 4th Nov'26 Conductor, a Mac app that runs several Claude Code, Codex and Cursor agents side by side in separate workspaces, announced it was joining Meta. Its changelog said Conductor Cloud would go read-only on 13th Nov and shut down on 20th Nov, with the desktop app getting updates until early 2027. On 13th Nov Conductor posted that it is not shutting down and has decided to keep developing the product, with regular updates resuming next week. It has not said why the Meta plan changed or whether Cloud workspaces are back to normal. Its earlier instructions for leaving still stand: a cloud workspace and its chats can be moved to the Mac with Transfer to local, and Settings, Account, Export data downloads a full archive.","theLine":"Judge an agent tool by what you can take with you when its owner's plans change.","teaser":"","source":{"name":"Conductor","url":"https://www.conductor.build/changelog/0.94.2-conductor-is-back"},"isLead":true,"isSharable":false,"keyFigure":"","figureLabel":""},{"id":"2026-11-13-claude-meetings-external-guests","url":"https://brewbytes.io/news/item/2026-11-13-claude-meetings-external-guests","headline":"Claude Meetings joins Google Meet calls and starts the work while people talk, but leaves when an outside guest joins","topic":"Tools","date":"2026-11-13","whatItMeans":"Meeting bots so far have produced a summary after the call, and someone still had to turn it into work. This one drafts the plan while the people who own it are still in the room to correct it, which is where most action lists go wrong.\n\nThe default on external guests is the part to copy if you build your own. A bot that can read your connected apps should not sit in a call with a client unless someone decided it should, and Anthropic made leaving the default rather than a setting someone has to remember. Before you switch it on, decide which apps it can reach, because anything it can read it can repeat to everyone on the call.","body":"On 10th Nov'26 Anthropic put Claude Meetings into beta for Team and Enterprise plans. Claude joins a Google Meet call as a visible participant, announces itself in the chat and works during the call, drafting the plan being discussed, mapping a timeline on a whiteboard, answering questions in the chat and pulling figures from connected work apps. Afterwards the notes, action items and anything Claude made are in claude.ai. By default it will not join a call with external guests unless an admin allows it, and it leaves if one joins. Google Meet is the only platform at launch, and Anthropic released Whiteboards and Sheets in beta alongside it.","theLine":"A meeting bot that can read your company's apps should leave when a client joins, and Claude's does by default.","teaser":"","source":{"name":"TechCrunch","url":"https://techcrunch.com/2026/11/10/anthropic-launches-claude-meetings-to-turn-workplace-discussions-into-plans-and-tasks/"},"isLead":true,"isSharable":false,"keyFigure":"","figureLabel":""},{"id":"2026-11-11-gemini-decisions-api","url":"https://brewbytes.io/news/item/2026-11-11-gemini-decisions-api","headline":"Google's new Decisions API answers yes, no or a score for $0.10 per million tokens, with no charge for output","topic":"Tools","date":"2026-11-11","whatItMeans":"A lot of what an agent spends on a general model goes on deciding rather than writing: which tool to call, whether a message breaks policy, which model to route to. Two of the big labs now sell a separate, cheaper model for exactly that, and it returns a confidence score you can set a threshold on, which a chat reply never gave you cleanly.\n\nLook through your pipeline for calls where you parse a yes or a label out of a written answer, and try one of them on a decision API against your own labelled examples. Google's speed and price figures are its own, so the test that counts is whether its confidence scores match how often it is right on your data.","body":"On 11th Nov'26 Google launched the Gemini Decisions API in Google AI Studio, running a new model called Gemini Decision 1. It answers only three kinds of question: yes or no, multiple choice, or a rating on a scale you set, and it returns each answer with a confidence score instead of written text. A request can carry up to 256,000 tokens of text, images, audio or video. Google prices it at $0.10 per million input tokens with output free, puts end-to-end latency at 150 to 250 milliseconds for contexts up to 16,000 tokens, and says it is nearly five times cheaper than Gemini 3.5 Flash-Lite. OpenAI launched its own Decisions API in October.","theLine":"Most of what an agent asks a model is a decision, and decisions now have their own cheaper model.","teaser":"","source":{"name":"SiliconANGLE","url":"https://siliconangle.com/2026/11/11/google-enters-the-decision-model-market-with-gemini-decisions-api/"},"isLead":false,"isSharable":true,"keyFigure":"$0.10","figureLabel":"per million input tokens on Google's Decisions API, output free"},{"id":"2026-11-11-united-bans-ai-agent-bookings","url":"https://brewbytes.io/news/item/2026-11-11-united-bans-ai-agent-bookings","headline":"United Airlines now bans AI agents from buying its tickets, even when the traveller approves each step","topic":"Policy","date":"2026-11-11","whatItMeans":"Most agent builders treat human approval as the thing that makes an automated purchase safe, and United has written into its contract that approval does not count. What the seller is protecting is its relationship with the customer and the upsells that come with it, so the risk to an agent that buys things is the merchant's terms rather than the law, and those terms can let the merchant cancel what your agent bought.\n\nIf your agent books, orders or pays on someone's behalf, check each merchant's terms for automated purchasing, and design the flow so the agent finds and fills the basket while the person completes the purchase on the merchant's own page. That handoff is what United still allows.","body":"On 11th Nov'26 United revised its contract of carriage to bar AI agents from making, changing or servicing a ticket, or buying any other service, and the rule applies even when the customer supervises or approves what the software does. It covers agents that come to United directly and those that book through travel agencies and aggregators, and a separate rule lets United cancel any reservation an agent made or changed. Agents can still search and show United's fares, and a person who researches with AI and then buys the ticket themselves is within the rules. Bookings agents completed before 11th Nov will be honoured. United cites exposed personal data, purchases the customer did not intend and miles spent without the passenger knowing, and its chief executive, Scott Kirby, also said he did not want another middleman. Delta has indicated that properly authenticated assistants can book, with extra safeguards for sensitive transactions.","theLine":"A traveller's approval makes an agent's purchase safe for the traveller, not acceptable to the seller.","teaser":"","source":{"name":"View from the Wing","url":"https://viewfromthewing.com/united-airlines-ceo-scott-kirby-bans-ai-agents-from-booking-tickets-doesnt-want-another-middleman/"},"isLead":true,"isSharable":false,"keyFigure":"","figureLabel":""},{"id":"2026-11-10-claude-fable-5-2-price-cut","url":"https://brewbytes.io/news/item/2026-11-10-claude-fable-5-2-price-cut","headline":"Anthropic cut Claude Fable 5.2's price by a fifth, but still tells most teams to default to the cheaper Opus 5.5","topic":"Models","date":"2026-11-10","whatItMeans":"The price cut matters less than where Fable now sits. At $8 and $40 it still costs twice as much as Opus 5.5, and Anthropic frames Fable as the advisor or orchestrator that plans while a cheaper model does most of the work. If you send every call to Fable today, route only the planning and the hard judgement calls to it and test whether Opus 5.5 holds on the rest.\n\nIf you build security or life-science tools, check the fallback setting before you switch. With it off, a blocked query on the API stays blocked; with it on, the answer comes from a different, older model, so log which model answered each request, because a result you trust from Fable may not have come from Fable.","body":"On 9th Nov'26 Anthropic released Claude Fable 5.2, its top general model, and Mythos 5.2, a version without some cyber and biology safeguards that only vetted organisations can use. Fable 5.2 costs $8 per million input tokens and $40 per million output, down from $10 and $50 for Fable 5.1, and it is on the Claude API as claude-fable-5-2 and on AWS, Google Cloud and Azure. Anthropic says it finishes hard tasks in fewer tokens than 5.1 and responds up to 30% faster. When its safety classifiers block a cyber or biology query, Anthropic's own apps answer from an older Opus model instead, and API users get that fallback only if they opt in. The same post recommends Opus 5.5, at $4 and $20, as the default for most coding and knowledge work. Fable 5.1 keeps its old price and has no retirement date.","theLine":"Fable got cheaper, but Anthropic's own advice is to let it plan and let a cheaper model do the work.","teaser":"","source":{"name":"Anthropic","url":"https://www.anthropic.com/claude-fable-and-mythos-5-2"},"isLead":true,"isSharable":false,"keyFigure":"20%","figureLabel":"cut to Claude Fable 5.2's API price"},{"id":"2026-11-09-openai-cursor-cutoff-cloud-agents","url":"https://brewbytes.io/news/item/2026-11-09-openai-cursor-cutoff-cloud-agents","headline":"OpenAI plans to pull its models from Cursor on 12th Nov, and your own API key only keeps them in the desktop app","topic":"Tools","date":"2026-11-09","whatItMeans":"Most people will read this as a model-picker change and paste in an API key, which fixes the part of Cursor a person sits in front of. The part that breaks is the work nobody is watching, such as a cloud agent that reviews pull requests overnight on a GPT model. Those agents run on models Cursor supplies, so once OpenAI's supply stops, any of them set to a GPT model needs a new one, and Cursor's new Apps are built by those same cloud agents.\n\nBefore 12th Nov'26, search your Cursor automations and pipeline scripts for any OpenAI model name, and decide for each whether it moves to another model in Cursor or out to Codex or the OpenAI API directly. If it changes model, rerun the checks you trusted it on, because a review agent that passed your tests on GPT has not yet passed them on anything else.","body":"OpenAI told SpaceX in late August, after SpaceX bought Cursor, that it intends to end the contract that supplies OpenAI models to Cursor, and proposed 12th Nov'26 as the cut-off. OpenAI's help page, last updated in late October, says the final date is still to be agreed between the two companies and that Cursor may end access sooner. OpenAI gives three ways to keep using its models inside Cursor: your own OpenAI API key, billed separately at API prices; the Codex extension, which runs as its own agent inside the editor; or a gateway such as Amazon Bedrock or Azure. All three cover only local Chat and Agent in the desktop app. Cursor's Tab completion, Auto model routing, Cloud and Background Agents, Automations, the Cursor CLI and its API and SDK cannot use your key or gateway credentials, because Cursor supplies the models for those itself. OpenAI's stated reason is that it is not confident SpaceX will use its technology within its terms.","theLine":"An API key keeps GPT in the editor you sit in; it does not reach the agents that run while you are away.","teaser":"","source":{"name":"OpenAI Help Center","url":"https://help.openai.com/en/articles/20001506-using-openai-models-in-cursor"},"isLead":true,"isSharable":false,"keyFigure":"12 Nov","figureLabel":"OpenAI's proposed cut-off for its models in Cursor"}]}