Gemini 4 Carbon Launch Pricing | AI Field Notes #134

AnalysisA chat window that can host Adobe's photo editor or a full Confluence page changes what a connector is. Anthropic (the AI lab behind Claude) said on November 10 that partner apps can open beside a conversation or full screen, with Claude making changes inside them once the user approves each one. Thirteen tools support it at launch, and Figma, Harvey and Shopify are building theirs. The plumbing is MCP Apps, an open extension to the Model Context Protocol (the standard that lets AI models call outside tools). Software companies now have to decide whether their front door is their own site or someone else's chat.

AnalysisGoogle's strongest model costs the same per token as its everyday one until November 30, then twice as much. Gemini 4 Carbon reached general availability on November 10 at $2 per million input tokens and $10 per million output, rising to $4 and $20 in December. Google says it tops CursorBench 4.0 (a coding test run by the editor maker Cursor) at under half the per-task cost of the next best outside model. The Decoder notes Carbon burns more tokens by design, and the testing firm Artificial Analysis put it at $7.76 per solved task against $6.25 for GPT-6 Astra Max. The per-token sticker tells you little about the invoice.

AnalysisLast year's models produced circuit boards that were about 95% complete and unusable, according to Diode Computers, which designs and builds them. On November 9 Anthropic said Claude Fable 5.2 designed three devices with Diode (a speakerphone, a voice hub and a handheld virtual pet), each with a schematic, board layout, enclosure and firmware. All three were manufactured straight from its files and worked with no circuitry fixes. When the standard routing tool proved too slow for the hub's 100-pin tracks, the model wrote its own router in about half an hour. Its score on Diode's layout test rose from 25% on Fable 5 to 69%. The prototypes have no certification yet.

AnalysisSeven in ten requirements met sounds fine until you count finished jobs. Co-Design Bench, built by researchers at Alibaba and several Asian universities and reported on November 8, hands coding agents real change requests for existing web apps, picked from sessions where users had to rephrase or roll back. GPT-6 Astra led ten models, meeting about 70% of requirements on average and getting 26.4% of tasks fully right. It solved under 18% of the repository tasks. Older tests are saturated above 99%, while the spread between models here passes 35 points. Rankings also shifted with the harness, the software that lets an agent read and run code.

AnalysisA browser tab was enough to steal a developer's secrets from one of the official demo servers for the Model Context Protocol (MCP, the standard that connects AI agents to tools). The server-everything package listened on every network interface with wildcard CORS (a setting that lets any website call it) and no login, and one of its tools returned the whole process environment. The security firm Adversa's November roundup also lists a TypeScript SDK bug in versions 2.0.0 through 2.2.0 that could send one client's responses to another, and a FastMCP flaw that hands a client secret to whichever server asks. A separate study of 21,097 MCP servers produced 13 advisories.

AnalysisThe company that pays people to teach AI their jobs is now worth $20 billion. Mercor announced a $750 million Series D on November 10, doubling the $10 billion valuation from its October 2025 round, with Nvidia and Salesforce among the backers. It says more than five million experts train and evaluate models through its marketplace. Co-CEO Brendan Foody said Mercor spends three times as much on inference (running AI models) as on employee salaries, and that "every job in the economy will need its own AI agents." Neither the funding statement nor Foody's post mentions the April breach that led Meta to freeze its data work with the company.

AnalysisOpenAI now shows companies how often ChatGPT mentions them, a number startups have been selling as an estimate. The Businesses in ChatGPT beta, announced November 10, gives verified firms a blue check and editable prices and stock levels, plus counts of mentions and click-throughs. Eligible US online sellers can also run a Business Agent that answers shoppers inside ChatGPT, labeled and kept apart from ChatGPT's own answers. An agent can be sponsored as the landing point for ads, a business that passed a $1 billion annual run rate in August with 100,000 advertisers. OpenAI says paying does not change what ChatGPT recommends. About 1.2 billion weekly users are relying on that.

AnalysisThe note-taker bot has been promoted to participant. Claude Meetings, announced November 10 in beta for Team and Enterprise plans, joins a Google Meet call under its own name and posts a notice in the chat. It answers questions there, pulls figures from connected work apps, and drafts the plan as it is discussed. Anthropic also shipped Whiteboards and Sheets, a shared planning board and a spreadsheet builder with formulas and dropdowns. Claude leaves a call automatically if an external guest joins, unless an admin has allowed otherwise. Microsoft and Google sell the office suites these tools sit on top of, and OpenAI launched its own workplace tools a few weeks ago.

AnalysisBuying generators before you need them turns out to carry a price. SoftBank reported a net loss of 1.12 trillion yen (about $7 billion) for its fiscal second quarter on November 10, including 597.2 billion yen, roughly $3.7 billion, from changed procurement plans for power generation equipment at US AI data center projects. It had paid $9.48 billion in advance for that equipment by the end of September because of shortage fears and long lead times. An Ohio campus with a 20-year, eight-gigawatt lease to OpenAI is among the sites. SoftBank carries its OpenAI stake at $99.6 billion against $54.6 billion invested, and its shares closed up 13.9% anyway.

AnalysisNvidia's stock lost more than 4% on Monday and another 5.47% on Tuesday, briefly dipping under a $4 trillion market value it first cleared in July 2025. On November 10 the board retained Joon H. Kim, a former acting US Attorney in Manhattan, and the law firm Wachtell Lipton to investigate CEO Jensen Huang and whether company secrets or national security were compromised. The board says it has no reason to believe they were. A shareholder class action landed in federal court in San Jose. Analyst Dan Ives puts the odds of Huang's removal near 15% and expects the probe to run four to six weeks. AMD rose 13% on Monday.

AnalysisStanford's once-every-five-years review of AI includes a takeaway most reports avoid: "This report will be out of date by the time you read it." The 2026 AI100 study, released November 10 by a 12-member panel chaired by Oxford's Michael Wooldridge, credits scaling (adding more compute and data) for most of the past five years' advances. It then says that approach is running short of fresh high-quality data while frontier compute grows prohibitively expensive, and expects progress to come from better algorithms and smaller models. On money the panel is blunt. AI's impact does not yet show up in macroeconomic data or company returns.