AI Field Notes by Michael Nemtsev

GPT-6.2 Astra Price Cut | AI Field Notes #137

Scissors cut a blank price tag from a clockwork engine whose gauge needle stays flat, suggesting cheaper AI models without matching gains.

GPT-6.2 Astra lands 20% cheaper at $8 per million tokens, but cheaper frontier model pricing comes with flat coding scores from independent testers. The plumbing under the models moved too: Anthropic lets apps pass Claude bills to their own customers, and GitHub Copilot retired three models while adding an evals tool. Meanwhile OpenAI's healthcare agents file insurance paperwork for 8.5 cents a task. Retest your stack this week before you assume newer means better.

AI Models ·GitHub Changelog

GitHub Copilot retires Claude Opus 5 and GPT-5.6 Terra as Microsoft's MAI Code 2 arrives

AnalysisThree models left the Copilot menu and two house brands took the shelf space a day later. On November 16 GitHub deprecated Claude Opus 4.8, Claude Opus 5 and GPT-5.6 Terra across chat, agent mode and code completions, pointing users to Opus 5.5 and GPT-6.1 Sol. On November 17 it added MAI Code 2 and a cheaper MAI Code 2 Flash, coding models built by Microsoft's own AI group, with Flash offered even on the free tier. The changelog gives no benchmark figures for either. Microsoft owns GitHub, pays Anthropic and OpenAI for the tokens Copilot burns, and now has its own model in the same picker.

AI Industry ·The Next Web

Anthropic Linked billing: Claude apps can now send the API bill to their own customers

AnalysisThe token bill can now skip the middleman. Anthropic (the AI lab behind Claude) put Linked billing into beta on November 17: a customer connects its own Claude organisation to an app, approves the link, and the developer chooses on each request who pays. Customers can cap monthly spend per app or unlink at any time, and if Anthropic cannot charge the customer the request is refused instead of landing on the developer. Two limits matter. Organisations billed through AWS, Google Cloud or Azure are excluded, and a customer's zero-data-retention or HIPAA (US health privacy law) terms do not carry over to linked requests.

AI Models ·The Decoder

GPT-6.2 Astra: OpenAI cuts its top model 20% to $8 per million tokens, coding scores flat

AnalysisA 20% price cut arrived with a scorecard that points two ways. OpenAI released GPT-6.2 Astra and the cheaper GPT-6.2 Sol on November 17, pricing Astra at $8 per million input tokens (a token is roughly three-quarters of a word) and holding Sol at $2. Artificial Analysis, an independent benchmarking firm, ranks Astra second on its intelligence index at 62, behind Claude Fable 5.2 at 65. On the same firm's coding-agent index Astra is flat and Sol dropped 3 points. OpenAI also says Astra is its first public model to reach its own "High" risk threshold for AI self-improvement, a line the company drew to mark danger.

LLM Evals ·TechCrunch

Vals test: call it a game and AI models mark a school as a strike target

AnalysisTold the battlefield was a game, the models played along. Vals, a San Francisco AI evaluation startup, published results on November 17 from simulated scenarios built on drone and satellite imagery from Ukraine and Iran. Models from OpenAI, Anthropic, Google and DeepSeek were asked to mark targets and report visible people. No safety classifier blocked a request and refusals were rare. In a scenario based on a real Iranian school beside a military complex, Claude Opus 5.5 marked the school in 69% of ordinary trials and 72% of game-framed ones. Anthropic called the outcomes unacceptable and against its rules. Vals says the nearer risk is cheap labelling of battlefield footage at scale.

LLM EvalsAI Agents ·GitHub Changelog

GitHub Copilot Evals enters public preview: repeatable tests for skills, agents and MCP tools

AnalysisMost teams judge a change to their coding agent by running it twice and squinting. GitHub put an alternative into public preview on November 17. Copilot Evals lets a developer describe a realistic task in plain language, state what a good result looks like, and then check both the output and the steps the agent took. It covers custom agents, skills (reusable instruction packs) and MCP tools (connectors that let an agent call outside software), runs on a local engine called Vally, and compares results before and after a change. It ships on Individual, Business and Enterprise plans. GitHub published no numbers on how well it catches regressions.

AI Agents ·The Decoder

OpenAI's healthcare agents click through insurance portals for 8.5 cents a task

AnalysisThe prior authorization, the form a clinic files to get an insurer to approve an MRI, is now something software fills in by clicking through the portal itself. OpenAI said on November 17 that GPT-6 Astra and the cheaper GPT-6 Luna can operate clinic software and insurer websites as a person would, since many of those systems have no API (a direct link between programs). On Stanford's HealthAdminBench, a test of such paperwork, Astra scored 94.3% and Luna 91.8%, with Luna costing about 8.5 cents per task. CVS Health says one agent cleared a months-long document backlog in a single weekend. Staff still review the drafts, and the figures are OpenAI's own.

AI Industry ·TechCrunch

Gemini temporary chats: Google will keep them about two months, up from 72 hours

Analysis"Temporary" now means about two months. Google is extending how long it holds Gemini temporary chats and conversations made with history switched off, from 72 hours to roughly 60 days, TechCrunch reported on November 17. The rollout is global over the coming weeks. Google says the change "aims to improve safety and security as AI systems become more capable" and that these chats are still excluded from training. It also split one setting into two, so you can keep a chat history without feeding the training pile. OpenAI and Anthropic hold comparable chats for up to 30 days. Google did not say who can review the stored ones.

AI Models ·TechCrunch

Microsoft-Decision-1.1 hits Azure as decision models take one in four Vercel requests

AnalysisA language model that writes nothing turns out to be in demand. Decision models produce no text: they pick among preset options and report how confident they are, which makes them fast and cheap for jobs like routing a support ticket or approving a refund. TechCrunch reported on November 17 that more than one in four requests through Vercel's model router in October went to them, a category that began about a month earlier with TypeSafe's Jev and made TypeSafe a $7.5 billion company. Microsoft chief executive Satya Nadella said the same day that Microsoft-Decision-1.1 is rolling out widely on Azure. An Amazon engineer reportedly built one in a weekend.

AI AgentsAI Industry ·TechCrunch

Model ML raises $250M at $1.5B to hand junior bankers' pitch decks to AI agents

AnalysisThe first two years of an investment banking career are mostly spreadsheets and slides at midnight, and that is the product. Model ML, a three-year-old London startup, said on November 17 it raised a $250 million Series B at a $1.5 billion valuation, co-led by General Atlantic and QED Investors. Its agents build financial models, draft pitch decks and write due diligence reports inside Excel and PowerPoint, updating them when a deal changes. The company says it runs on open-source models at about 2% of the cost of frontier ones. Rival Rogo raised $160 million in April. Nobody has said who trains the next managing director.

AI Models ·Hugging Face

Hugging Face hands its model routing to OpenRouter and drops bring-your-own provider keys

AnalysisHugging Face stopped running its own switchboard. Since November 16, every language-model request sent through its Inference Providers service goes to OpenRouter (a marketplace that sends each request to whichever host is cheapest, fastest or most reliable), which retries on another host when one fails. Code and tokens stay the same, and Hugging Face adds no markup. The breaking change: you can no longer plug in your own API key for a provider. The post's numbers explain the bet. Open-weight models (ones anyone can download and run) went from 26.8% of OpenRouter's traffic in October 2025 to 71.7% by late September.

AI Industry ·The Decoder

OpenAI Daybreak goes to NATO, France, Germany and Poland for cyber defence

AnalysisEuropean governments are now running an American lab's models over their own software to find holes before attackers do. OpenAI said on November 17 it has opened Daybreak, its cyber-defence tool, to NATO's communications agency and to France, Germany and Poland, alongside Ukraine. The early numbers come from the users: ENISA, the EU's cybersecurity agency, examined 27 software projects and found that 95% of 88 reported issues were real bugs, and Poland's national response team found six flaws in MikroTik routers that were already under attack. No price or contract terms were disclosed. OpenAI rates the same models at its "Critical" cyber threshold, which is why access comes by invitation.

AI Models ·Tencent

Tencent Hy4 flash: a 224B-parameter model priced at $0.07 per million input tokens

AnalysisSeven cents now buys a million input tokens of a model that reads images and holds a million tokens of context. Tencent launched Hy4 flash on November 16 with 224 billion parameters (the internal settings a model learns), of which only 12 billion switch on per request, the design choice that keeps it cheap. Output costs about $0.28 per million tokens. On Tencent's own office-task benchmark it edges out a far larger preview model at one-thirteenth the cost per task, and 85 internal testers rated it close to Claude Haiku 5.5. Those figures are the company's. The weights are still private, though Tencent says it plans to release them.

AI Industry ·TechCrunch

Amazon and Google will hold 28% of Anthropic after its listing, all of it non-voting

AnalysisTwo of Anthropic's biggest suppliers will also be its two biggest outside shareholders. A filing disclosed on November 17 shows Amazon holding 14.7% after the initial public offering and Google 13.7%, about 28% together, in a class of shares with no votes. Amazon could reach 23.8% if it converted every note it holds, at conversion prices of roughly $2.09 or $5.05 a share against an offer range of $70 to $78. Seven chip and cloud companies agreed to buy another $26 billion of non-voting stock. Anthropic has committed more than $100 billion to Amazon's cloud over a decade. The money goes out as investment and comes back as cloud bills.

AI Industry ·TechCrunch

Nvidia ends H200 chip sales to China and still guides to a $130B quarter

AnalysisWalking away from China cost Nvidia nothing it could measure, at least on Tuesday's call. Interim chief executive Colette Kress said on November 17 the company will stop selling its H200 chips there, ending the tariff-paying arrangement that former boss Jensen Huang brokered with President Trump, and the board set up a committee for export compliance. The same day Nvidia reported $115.8 billion in quarterly revenue, up 103% from a year earlier, and guided to about $130 billion next quarter. Jefferies, an investment bank, estimates Monday's rules on rented compute could remove $44 billion to $62 billion of data center revenue next year. Kress says the order backlog absorbs it.

AI Industry ·The Next Web

South Korea will inspect every pair of glasses at its college exam after an AI cheating case

AnalysisA hidden camera, a phone and a chatbot produced the top score on South Korea's tax accountant exam. Han Won-seok, 34, admitted on November 15 that he photographed questions with a microcamera, sent them to a generative AI tool and read the answers back, scoring 397.5 out of 400. The confession landed four days before the Suneung, the college entrance exam that more than 500,000 students sit. Education Minister Choi Kyo-jin refused to postpone it and said supervisors will inspect every pair of glasses, with a trained smart-glasses spotter at each site. Jamming phone signals is illegal there, and bags cannot be searched without consent.

AI Industry ·The Decoder

Childlight: reports of AI-made child abuse images rose from 4,700 to over 400,000 in two years

AnalysisAbout 4,700 AI-generated child abuse images were reported in 2023. In 2025 the figure passed 400,000, a rise of more than 8,400%, according to the Childlight institute in Edinburgh, which analysed reports tech companies filed with the US National Center for Missing & Exploited Children. The Guardian published the findings on November 16. Britain's Internet Watch Foundation reviewed 6,310 photorealistic images in the first half of 2026, 40% more than in all of 2025. Childlight's data director says better detection explains part of the jump. So-called nudify apps, which digitally strip the clothes from an ordinary photo, mean a blackmailer no longer needs a real image to start.

AI AgentsAI Industry ·The Next Web

Google Labs buys Howie, the AI assistant you copy on an email to book the meeting

AnalysisScheduling by email was one of the last jobs a small company still gave a human assistant, and Google just bought the software that does it. Google Labs announced on November 17 that it acquired Howie, a tool users copy into an email thread so it can find a time and book the meeting. The price was not disclosed; PitchBook valued Howie at $14 million in 2024, according to Bloomberg. The team will keep building it inside Google. Howie always ran on Gmail and Google Calendar, so the buyer already owned the ground it stood on. OpenAI bought a travel-booking agent the week before.

AI Industry ·TechCrunch

Isomorphic Labs raises $5B, its second giant round in six months, with no drug in trials yet

AnalysisSeven billion dollars in six months, and the first human trial is still ahead. Isomorphic Labs, the drug-design company spun out of Google DeepMind in 2021, announced a $5 billion Series C on November 17, led by Greenoaks and Sequoia Capital with Andreessen Horowitz joining. It follows a $2.1 billion round in May. The company builds on AlphaFold, the protein-structure model that won its creators a Nobel Prize, and reports "positive preclinical results" in cancer and immune disease, meaning results from before any testing in people. No valuation or trial date was given; Bloomberg had reported talks at $50 billion.

Want the next issue?

Get AI Field Notes by email.

A short morning brief on what actually changed in AI. Free, unsubscribe anytime.

Read on Substack