Welcome to Blank Metal’s Weekly AI Headlines.
Each week, our team shares the AI stories that caught our attention: the articles, announcements, and insights we’re actually discussing internally. We curate the best of what we’re reading and add the context that matters: what happened, why it matters, and what to do about it.
Investors Put a Price on AI
Two of the companies most AI plans depend on reported to public investors this week. Anthropic priced its offering above the range it set a week ago and closed its first day 17% above that price, and Nvidia posted its first results since its founder resigned. For buyers, both bear on the price and terms of what you buy next year.
Anthropic Prices Its IPO at $83 a Share and Closes Its First Day Up 17%
What: Anthropic priced its initial public offering on November 18 at $83 per share, above the $70 to $78 range it set when the roadshow began. The offering covers 980 million Class A shares, about 956 million sold by the company and about 24 million by existing stockholders, and the underwriters have a 30-day option on another 147 million. Reuters put the proceeds at $81 billion, a record for an IPO. Shares began trading on the Nasdaq on November 19 under the symbol ANT and opened at $111.50, 34% above the offer price, then gave back about half of that gain to close at $97.17, up 17%. Reuters put the valuation reached during the session at about $2 trillion counting employee stock options and unvested shares, and at as much as $1.5 trillion on shares outstanding. Reuters also reported a separate $26 billion private placement with AMD, Broadcom, Google, Micron, Microsoft, Nvidia, and TSMC, nearly $900 billion in computing commitments with partners including Amazon and Google, and revenue growth of more than 1,100% in the quarter ended September 30 from a year earlier.
So What: One of your model vendors will now file quarterly reports. Its revenue, margins, and compute obligations become public record, and the numbers already show the scale: close to $900 billion in computing commitments, and a $26 billion placement whose buyers include its own chip and cloud suppliers. The company also answers to shareholders who paid a price that assumes fast growth, and that pressure usually reaches customers through pricing and minimum commitments.
Now What: When Anthropic’s first quarterly report lands, read the sections on compute commitments and customer concentration. If a multi-year Claude agreement is still open, close price protection and capacity terms before that first earnings cycle. Check whether your contracts with any AI vendor let you move workloads to a second model without a penalty, and test that path once on a real workload.
Nvidia Reports $115.8 Billion Quarter; Chair Says Jensen Huang Will Not Return as CEO
What: Nvidia reported third-quarter revenue of $115.8 billion on November 17, up 20% from the prior quarter and 103% from a year earlier, with a GAAP gross margin of 74.8%. It was the first report since Jensen Huang resigned on November 12. “Vera Rubin is off to an explosive start, our roadmap is on track and our ambition has never been greater,” said Colette Kress, the interim chief executive and chief financial officer. Reuters reported that Nvidia forecast fourth-quarter revenue of $130 billion, plus or minus 3%, above the $125.25 billion analysts expected. Board chair Harvey Jones opened the investor call by saying the company would not rehire Huang as CEO and would not take questions about his departure, according to Reuters. The board has hired the search firm Heidrick & Struggles. “We want to install the CEO, the permanent CEO, as quickly as possible, but we won’t expedite the CEO before finding exactly the right person,” Jones said. Kress told analysts that Nvidia has made no changes to how it backs its commitments to AI cloud providers, and, according to Reuters, the board created a committee on national security and export controls the same day. Shares rose 4% in extended trading.
So What: Last week’s open questions were whether Nvidia’s customer financing would continue and who would set the roadmap. The first has an answer for now: Nvidia says the financing that helps smaller AI cloud providers buy chips is unchanged. The second is still open, with a search under way and no timeline. The results and the forecast point to demand that has not slowed, which keeps capacity tight and expensive for anyone buying it.
Now What: Teams holding GPU reservations with a smaller AI cloud provider should ask what in the contract protects them if Nvidia’s backing changes under a new CEO. Keep infrastructure plans where they are. If you’re budgeting 2027 compute, assume prices hold at current levels, and price out at least one option that runs on a different chip family.
Three Vendors Cut the Cost of Capable Models
OpenAI, Google, and Microsoft each released models this week, and each led with price. OpenAI matched Anthropic’s top-tier rate, Google released a model at well under half that rate that it says places first or second on several benchmarks, and Microsoft shipped specialized coding and security models it says cost a small fraction of general-purpose ones. Each cost comparison below is the vendor’s own.
OpenAI Releases GPT-6.2 Astra and Sol and Cuts the Price of Its Top Model by 20%
What: OpenAI released GPT-6.2 Astra and GPT-6.2 Sol on November 17. OpenAI positions Astra for the hardest work and Sol for everyday work. Astra costs $8 per million input tokens and $40 per million output tokens, 20% below GPT-6 Astra, with cached input down 60% to $0.40. Sol stays at $2 and $10. Developer Simon Willison noted that Astra’s new price is identical to Claude Fable 5.2’s. Both models are available in ChatGPT Work, Codex, and the API; Enterprise and Edu users get access once a workspace admin turns it on. OpenAI says Astra scores 5.5 points above Claude Fable 5.2 on the Agents’ Last Exam benchmark “at just 5% of the cost” per task. OpenAI classifies both models at its Critical threshold for cybersecurity capability, and says Astra is its first externally released model to meet its High threshold for AI self-improvement, “which measures a model’s ability to accelerate AI research.” The models ship with tighter restrictions on some dual-use cybersecurity and biology requests. Willison described early reaction as mixed, writing that Astra “isn’t scoring quite as high as Fable 5.2 on various benchmarks,” and quoted one pre-release reviewer who found that on long tasks the model stops to ask about internal details the user has no way to answer.
So What: OpenAI’s and Anthropic’s top models now carry the same list price, so the comparison that matters is cost per finished task, and OpenAI’s claim of a twentyfold advantage on one benchmark is a number to test on your own work. The reviewer’s report of a model that stops mid-job to ask unanswerable questions bears on how much supervision an unattended run needs. The new safeguards may also block legitimate security testing your team already does.
Now What: If you run GPT-6 Astra in production, rerun your evaluation set on 6.2 and compare cost per completed task, including the cached-input discount if your agents run long sessions. Test one long unattended job and count how often it stops for input. If your security team uses OpenAI models for vulnerability research, check which workflows the new restrictions block and ask OpenAI what access keeps them running. Enterprise admins need to turn the models on before users see them.
Google Releases Gemini 4 Neon at $3 per Million Input Tokens
What: Google released Gemini 4 Neon on November 18. Google places it between Gemini 4 Carbon, its model for the hardest long-running projects, and Gemini 4 Hydrogen, its model for fast everyday tasks, and describes it as “built for complex coding, analytical reasoning, and production-grade enterprise workflows.” It costs $3 per million input tokens and $15 per million output tokens, with cached input at $0.15, and has a 1M-token context window. Among the results Google cites: second place on the Vals Index at 70.9%, first place on Zapier’s AutomationBench business workflow test at 56.3%, and 46.2% on Cognition’s FrontierCode 2.0 engineering test at its default setting. Glean, an enterprise search company, said that in its early evaluations Neon “matched or exceeded Gemini 4 Carbon on overall quality at 3x the speed and 60% of the cost.” Neon is available through the Gemini API, Gemini for business, Google Antigravity, and the Gemini app for paid subscribers. A version for security work is rolling out to members of Google’s Fairwind Program for vetted defenders.
So What: By Google’s account, a model priced at well under half the top-tier rate places first or second on benchmarks run by outside firms, one of them a business workflow test. For most enterprise work (document analysis, routine coding, multi-step workflows), the middle tier may now be enough. The top tier’s remaining case is the hardest, longest-running jobs, and for many companies that may be a smaller share of usage than their spending suggests.
Now What: Pull last month’s usage and sort it by task type. For each type running on a top-tier model, test a mid-tier model (Neon, or the equivalent from your current vendor) on a sample of real work and compare quality, speed, and cost per finished task. Move what passes. Keep the top tier for work where your own tests show a difference.
Microsoft Introduces Six of Its Own AI Models, Led by a Coding Model for GitHub Copilot
What: Microsoft announced six AI models of its own at its Ignite conference on November 17, covering coding, cybersecurity, image generation, and voice. Axios reported Microsoft’s claims for them: MAI Code 2 delivers 98% of GPT-6.1 Sol’s performance at 40% of the price and 96% of Claude Opus 5.5’s at 5% of the cost, and MAI Cyber 2 Flash beats GPT-6 Astra on a security benchmark at less than 2% of the cost per task. Axios noted that the comparisons “rely on selected benchmarks, and some pricing advantages reflect promotional discounts.” Chief executive Satya Nadella said Code 2 and the smaller Code 2 Flash are generally available in GitHub Copilot and in Microsoft Foundry. Axios noted that a model tuned for everyday coding won’t be as capable at advanced math or scientific research. Microsoft AI chief Mustafa Suleyman told Axios the company is also building a general-purpose frontier model and declined to give a release timeline. Enterprise customers can license the model weights to customize with their own data, provided they deploy within Azure.
So What: Microsoft has relied on OpenAI and Anthropic for its best models, and it now has its own in the product where many enterprises do their coding. If you buy GitHub Copilot, Microsoft’s own models are now among the ones that can sit behind a request, so it is worth knowing which one does. Specialized models that do one job at a fraction of the cost are a reasonable trade for routine work. The trade is narrower capability, and Microsoft picked the benchmarks behind its numbers.
Now What: For teams on GitHub Copilot: find out which model handles which requests today and whether your admin settings let you choose. Run Code 2 against your current default on a sample of your own pull requests before changing anything, and check what the price is after any promotional period ends. If you have a security operations team, ask for the benchmark behind the Cyber 2 Flash claim and test it on your own alert data.
The Assistant Joins the Workday
Microsoft and Anthropic both moved their assistants further into ordinary office work this week, into documents, meetings, and the follow-up work after them. Both arrive with new admin controls, and the rollout differs by plan or program.
Microsoft Uses Ignite to Put Office, an Always-On Agent, and App Building Inside Copilot
What: At Ignite on November 17, Microsoft announced a set of changes to Microsoft 365 Copilot. Home, a starting screen that combines chat and Copilot’s Cowork agent, and Office in Copilot, which lets people create and co-edit Word, Excel, and PowerPoint files inside the Copilot app, are generally available. Autopilot, an agent that works without being prompted (watching Teams channels, following up on email, running recurring work), began rolling out to customers in Microsoft’s Frontier early-access program. Each Autopilot has its own identity in Microsoft Entra and a named manager, IT can limit or block its access, and its sign-ins and actions are logged under its own name. Copilot Code, which builds apps and dashboards from a plain-language description, will be generally available in the coming weeks, Microsoft says, and the apps it produces are hosted in the customer’s tenant and listed in the Microsoft 365 admin center. Employees can choose models from OpenAI or Anthropic where company policy allows. Microsoft also expanded Agent 365 with cost management: spending by user, agent, and model, with limits and alerts. Nadella said Agent 365 already tracks more than 50 million agents.
So What: Microsoft is treating agents as a class of worker that IT has to manage, with an identity, a manager, an audit trail, and a budget. That framing is useful whatever you buy, because it names the controls any always-on agent needs. Copilot Code also means employees outside engineering will start building small apps on company data, which helps where IT has a backlog and creates a new inventory to track.
Now What: Decide who owns agent identity and agent spending before Autopilot reaches your Microsoft 365 tenant, and set spending limits as soon as the cost controls are available to you. Set a simple policy for employee-built apps: who can publish, what data they can touch, and who reviews them. If you use another vendor’s agents, check whether you can answer the same four questions for each one: who it is, who manages it, what it did, and what it cost.
Claude Can Now Join Video Calls and Start the Work While the Meeting Is Still Going
What: Anthropic began rolling out Claude Meetings in beta on November 12 for Team and Enterprise plans. Claude joins a video call as a visible participant, follows the conversation, answers questions aloud or in chat, and builds documents, whiteboards, and other files while people talk. Afterward it saves notes, action items, a transcript, a summary, and descriptions of what was shown on screen; it does not save audio. Google Meet is the only platform supported at launch. “When Claude joins, it appears as a participant and posts a notice in the call’s chat,” Anthropic says, and the person adding it is responsible for making sure everyone on the call is fine with being transcribed. By default Claude stays out of calls that include anyone from outside the organization and leaves if an outside guest joins; an admin can change that. On Team plans the feature is on by default. On Enterprise plans an admin has to turn it on. During the beta it isn’t available to organizations using zero data retention, HIPAA-ready configurations, or customer-managed encryption keys. Notes and transcripts are free during the beta.
So What: Unlike most note-takers, this assistant can act on what it hears, using the same connected tools it has in chat, so a decision made early in a call can be a draft document before it ends. It also means meeting transcripts become part of what your AI vendor stores, and the default on Team plans is that any employee can invite it.
Now What: Claude Team plans have this on by default. Check this week whether that fits your recording and retention policies, and tell employees what the rule is for calls with customers or candidates. Check your state and country consent requirements for transcription before allowing it on calls with outside guests. Then pick one recurring internal meeting, try it for a month, and judge it on whether follow-up work gets done faster.
How Far to Trust an Agent
Three stories this week turn on what has to be true before you rely on work an accountable vendor or your own staff didn’t do. For payments, OpenAI’s answer is an approval on every purchase. For a self-hosted copy of a paid tool, the answer sits in the fine print. For code, Anthropic’s Claude Code team puts its effort into the checks.
OpenAI Launches ChatGPT Wallet So Its Agents Can Pay Without Seeing a Card Number
What: OpenAI introduced ChatGPT Wallet on November 13. It lets agents in ChatGPT Work, Codex, and OpenAI’s “dots” product complete purchases on a user’s behalf. The user adds a card once. Before each purchase the user reviews the items, the total, and the card, then approves with a security check such as Face ID or a fingerprint. The payment network then issues a payment token, a substitute for the real card details, so neither the agent nor the business receives the actual card number. Users keep their card’s rewards and dispute protections. OpenAI named American Express, Chase, Mastercard, and Visa as partners, along with payment processors Adyen, Checkout.com, and Stripe and fraud prevention firms Forter, Riskified, Sardine, and Signifyd. The wallet is rolling out to ChatGPT Pro subscribers aged 18 and older in the US with eligible US-issued cards. OpenAI says it expects users will later be able to pre-approve purchases within set spending limits, categories, or time periods. “An agent making a purchase should strengthen, not bypass, the trust built into payments,” said Sherri Haymond, Mastercard’s global head of digital commercialization, in OpenAI’s announcement.
So What: OpenAI has the major card networks and several large processors behind one way for an agent to pay: a token, a per-purchase approval, and a record of what the customer authorized. If you sell online, some of your orders will start arriving this way, and OpenAI’s own guidance warns that fraud systems may decline legitimate agent purchases and that site security tools may block agents before they reach checkout. Orders lost that way won’t show up as declines in most reports.
Now What: Ask your payments and fraud providers how they identify agent-initiated purchases and what their decline rate on them is. Have your web team confirm that your CDN and bot protection recognize signed agent requests (the standard OpenAI names is Web Bot Auth), so you aren’t blocking buyers by accident. Make sure your product feed carries current prices and availability, since that is what an agent reads.
An Investor and an AI Coding Company Each Publish Their Own Version of E-Signature Software
What: On November 16, investor Kevin Rose released SignuDoc, an open source document signing application that a company deploys to its own cloud account. In his announcement he wrote that DocuSign “wants $8,100/yr from a 15-person team just to sign PDFs.” The project’s documentation says it covers sending documents to up to 10 signers, templates, reminders, identity checks by access code or emailed one-time code, a tamper-evident seal, and an audit trail, and that it runs at no cost for about 250 documents a month on free hosting tiers. The documentation is also specific about limits. It says SignuDoc is not affiliated with DocuSign, comes with no warranty, “does not give legal advice,” and is not built for health information covered by HIPAA or for documents the law keeps out of electronic signing, such as wills. Two days later, Replit said on X that its Missions agent product had crawled more than 1,000 flows in what it called the leading enterprise e-sign platform and “rebuilt them 1:1” as an application named Ternsign. Replit’s description of the result has not been independently reviewed.
So What: The cost of building a working copy of a common business application has fallen far enough that individuals and vendors are doing it and publishing the result. For a buyer, that changes the conversation at renewal for simple, per-seat tools. It doesn’t remove what the subscription also pays for: someone accountable for security patches, uptime, legal acceptance, and support. SignuDoc’s own documentation lists what it won’t cover, and a company that deploys it takes on those obligations.
Now What: List the per-seat tools you pay for where your use is simple and the data is low-risk. For one of them, estimate what it would cost to run an open source or agent-built alternative, counting the engineer who patches it and the counsel who signs off on it. Use that number in your next renewal negotiation, and build the replacement only where the math holds after those costs. Leave tools that carry legal, regulatory, or customer-facing risk with a vendor who is accountable for them.
How Anthropic’s Claude Code Team Checks Code That Agents Write and Review
What: Anthropic published a description on November 16 of how its Claude Code team verifies software, written by Addy Osmani. “On the Claude Code team, the vast majority of our code (~90%) is written and reviewed by Claude,” the post says. The team’s approach is to have agents prove their work with evidence: end-to-end tests that exercise the running product the way a user would, property tests that check rules across many generated inputs, and differential tests that compare new behavior against old. In Anthropic’s internal testing, a separate grader checking Claude’s work against a rubric “improved task success by up to 10 points” over a standard loop. The post spends as much time on keeping the checks honest. A code change cannot edit the review rules that judge it. Claude may not skip a flaky test, add delays, or retry until it passes. “If a check is still failing after three rounds, Claude stops and brings in a human.” The post also recommends running a new automated check in observe-only mode before letting it block work.
So What: A team that lets agents write and review nine-tenths of its code has moved its engineers’ attention to defining what done means and to maintaining the checks. The specific rules are the useful part, because they describe how automated checking fails in practice: an agent weakens the test, retries until it passes, or edits the rule that judges it.
Now What: Before raising how much code your agents write, look at what stands between their output and production. Confirm that an agent cannot modify the tests or review rules that evaluate its own change. Decide what happens to flaky tests, because an agent will learn to rerun them. Set a limit on automated retries before a person is pulled in. If your team has few end-to-end tests, start there, since the post names them as the one type to pick if you can only pick one.
Blank Metal is an AI consulting and engineering firm. We help organizations move from AI experiments to production systems. Learn more



