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.
Two Ends of the Price List
Anthropic and Meta released models on the same day at opposite ends of the market. One is priced for the hardest work a company has, the other for the small steps agents repeat thousands of times a day. Agent systems in production increasingly mix both, so the question for buyers is which work goes where.
Anthropic Releases Claude Fable 5.2 and Mythos 5.2 at a Lower Price
What: Anthropic released Claude Fable 5.2 and Claude Mythos 5.2 on November 9. The two are the same underlying model. Fable carries cybersecurity and biology safeguards and is available to everyone; Mythos is open to organizations approved through Anthropic’s Cyber and Life Sciences Verification Programs. Anthropic says the models generate output up to 30% faster than the 5.1 versions at a lower price, $8 per million input tokens and $40 per million output tokens, with an optional fast mode at double that. Fable 5.2 is available on the Claude Platform, Amazon Web Services, Google Cloud, and Microsoft Azure. On Anthropic’s published benchmarks, Fable 5.2 scores 69.9% on the FrontierSWE v2 coding test, against 56.3% for Fable 5.1 and 65.5% for GPT-6 Astra, and trails Gemini 4 Carbon on the AutomationBench business workflow test, 51.4% to 53.9%. Anthropic also says it took steps during training to limit gains in the research skills that could most directly speed up AI development, and that both models voice suspicion that they are being tested more often than any of its recent models. In a review published the same day, Wharton professor Ethan Mollick found that the model’s built-in reviewers caught factual errors, but “none of them asked whether the work was too long, too repetitive, or no fun.”
So What: Anthropic says its top model is now cheaper and up to 30% faster, which changes the math on work you may have ruled out on cost three months ago. The benchmark table is mixed enough to show that no single vendor leads every category. Mollick’s review points at the part that still needs a person: in his tests, the model fixed what it was told was wrong and didn’t flag output that was bloated or off the point.
Now What: If you run Fable 5.1 or Opus 5.5 in production, rerun your own evaluation set on 5.2 before switching, and compare cost per finished task, not cost per token. For long unattended runs, assign a named reviewer whose job is to judge whether the result is useful and the right length, since in Mollick’s review the built-in checks caught errors but not length or usefulness. If your security or research teams need the unrestricted model, start the verification application now.
Meta Releases Muse Ember, a Low-Cost Model for the Repetitive Steps Agents Run
What: Meta released Muse Ember on November 9, describing it as its smallest and most efficient model, built for “the work agents repeat all day: tool calls, coding subtasks, routing, extraction, and long document parsing.” It costs $0.10 per million input tokens and $0.50 per million output tokens, has a 1M-token context window, and accepts text, images, audio, and video. It is available through the Meta Model API, in Muse Code, and on Google Cloud and Oracle Cloud. Meta cites Artificial Analysis data showing Ember matches Claude Haiku 5.5’s best score on that firm’s intelligence index at about a third of the cost per task; on the FrontierCode 2.0 engineering test in Meta’s own table, Haiku 5.5 scores higher, 34.9 to 27.2. Developer Simon Willison notes that list prices for Ember, Haiku 5.5, and GPT-6 Luna are identical on short prompts and diverge on long ones (Ember triples above 200,000 tokens), and that Fable 5.2 costs 80 times as much per token.
So What: Three vendors now sell a small model at the same short-prompt list price, so the choice comes down to how each one performs on your tasks and what it costs when prompts get long. The gap between the small tier and the top tier is 80 times per token. An agent system that sends every step to the top model is paying that multiple on steps a small model may handle just as well.
Now What: If you run agents in production, break your token spend down by step type and find the routine steps (routing, extraction, simple tool calls) that could move to a small model. Test two or three small models on those steps with your own data, and price the test at your real prompt lengths, since the long-prompt surcharges differ by vendor.
The Companies Behind the Compute
Two of the companies that most AI plans depend on went through a leadership change and the start of a public listing this week. Nvidia’s founder and chief executive resigned, with an interim CEO in place and a search under way, and Anthropic opened the roadshow for an offering that could value it at about $1.4 trillion. For buyers, both bear on how much of a roadmap rests on one supplier.
Nvidia CEO Jensen Huang Resigns; CFO Colette Kress Named Interim CEO
What: Nvidia’s board announced on November 12 that Jensen Huang has resigned as president, chief executive, and a board member, effective immediately. Chief financial officer Colette Kress will serve as interim CEO while keeping the CFO role, and the board has started a search that will consider internal and external candidates. Harvey Jones, a director since 1993, was named the board’s first chair. The board said it expects “operations, customer commitments and product roadmap to continue without interruption.” The board’s announcement did not give a reason. The Associated Press reported that the resignation came after reports of a relationship between Huang and a former executive of Megaspeed, a Singapore-based Nvidia customer, and that the former executive is suspected of being a Chinese government agent. According to AP, Nvidia has said it has no reason to believe national security or its secrets were compromised. TechCrunch reported that shares fell nearly 12% in early trading, and that as of its second quarter Nvidia held $99 billion in investments and $36 billion in cloud service commitments that help AI cloud operators finance the chips they buy. Nvidia reports quarterly results on November 17.
So What: Nvidia supplies much of the compute behind the models and cloud capacity you buy, and the investments and cloud commitments TechCrunch describes have helped AI cloud operators pay for it. A change at the top doesn’t alter this quarter’s shipments. The open questions are slower ones, starting with whether that financing continues on the same terms and who sets the product roadmap. Those reach buyers through the price and availability of capacity over the next year or two.
Now What: Teams with GPU reservations or contracts at smaller AI cloud providers should ask how much of their capacity depends on Nvidia financing or purchase commitments, and what happens to a reservation if those terms change. Listen to the November 17 earnings call for what the interim CEO says about supply commitments and customer financing. Hold off on changing infrastructure plans until there’s more than a headline to go on.
Anthropic Opens Its IPO Roadshow at a Valuation of Up to $1.4 Trillion
What: Anthropic launched the roadshow for its initial public offering on November 10. An updated filing says it will offer 825 million shares at $70 to $78 each and has applied to list on the Nasdaq under the symbol ANT. CNBC reported that at the top of the range the offering would raise up to $64.35 billion and value the company at up to $1.405 trillion on a fully diluted basis, up from $965 billion in May, and that the company plans to price on November 18 and begin trading November 19, citing people familiar with the details. In the roadshow video, as transcribed by developer Simon Willison, co-founder Tom Brown says Anthropic trains and serves Claude across four chip families (AWS Trainium, Google TPUs, Nvidia GPUs, and AMD GPUs).
So What: Once Anthropic is public, its revenue, margins, compute commitments, and risk disclosures arrive every quarter, which gives customers far better material for vendor risk reviews than a private company provides. Running on four chip families means less exposure to trouble at any one supplier. Public shareholders will also expect growth, and that tends to show up in pricing and in pressure to commit to more volume.
Now What: If a multi-year Claude commitment is in negotiation, settle price protection and capacity terms now, before the first earnings cycle sets new targets. Ask your model vendor or cloud provider which chip platforms serve your workloads and whether you can specify them. Then add Anthropic’s quarterly filings to your vendor review calendar, starting with the first report after the listing.
Work Moves Inside the Assistant
Anthropic and OpenAI both spent the week pulling other companies’ products into their own. Anthropic put partner applications, spreadsheets, and whiteboards inside Claude, and OpenAI gave businesses a managed presence inside ChatGPT. Both companies want the assistant to be where work and buying decisions happen.
Claude Can Now Open Partner Apps Full Screen and Work in Them, and Adds Its Own Sheets and Whiteboards
What: On November 10, Anthropic said that connected tools can now open beside a Claude conversation or full screen, and that Claude can make changes directly in an open page, row, or file, with the user approving its actions. More than a dozen tools are supported at launch, including Asana, Atlassian, Box, HubSpot, and Notion; Figma, Harvey, and Shopify are building integrations. The feature runs on MCP Apps, an open extension to the Model Context Protocol, and Anthropic says an app built this way can run in any AI application that supports the extension. The same day Anthropic introduced Claude Sheets (spreadsheets with tabs, formulas, and tracked edits, exportable to .xlsx or CSV), Claude Whiteboards (shared boards for team brainstorming), and additions to Claude Data including certified metric definitions and live-shared dashboards. These are available on all paid plans. On Enterprise plans, Sheets and Whiteboards are off by default, and Claude Data turns on by default on November 19. Anthropic says Sheets has “simple spreadsheet functionality” for now.
So What: Anthropic is positioning Claude as the screen where people open their other tools, and Claude now has basic versions of documents, spreadsheets, and whiteboards of its own. That raises two practical issues for an IT or operations lead. An agent that can edit a live Confluence page or CRM record acts with the user’s permissions, so connector settings now govern real changes to systems of record. And some of what you pay for in separate collaboration tools is starting to overlap with what your AI subscription includes.
Now What: Admins on a Claude Team or Enterprise plan should review which connectors are enabled and which of their tools are allowed before the new features reach users, and decide whether Claude Data should stay on when the default changes November 19. Pick one team to try Sheets and Whiteboards on real work and report back on what they can and can’t replace. At your next collaboration software renewal, check for overlap before you add seats.
OpenAI Lets Businesses Claim a Profile in ChatGPT and Run Their Own Agent There
What: OpenAI introduced Businesses in ChatGPT in beta on November 10. Through a new Business Hub, a company can claim and verify its profile, update its details, add product feeds with pricing and availability, and see aggregated data on how often it is mentioned, how many clicks it receives, and the topics of conversations where its website is cited. OpenAI says ChatGPT has 1.2 billion weekly users. Eligible ecommerce and online service businesses in the US can also build a Business Agent, which answers customer questions in a conversation that is labeled and kept separate from ChatGPT’s own answers. Businesses can sponsor their agent and use it as the destination for ChatGPT ads. OpenAI says it does not share users’ conversations or Business Agent chats with businesses, and that insights appear only above minimum volume thresholds.
So What: Customers already ask ChatGPT which product to buy, and until now a company had little say in what it answered about them and no data on how often they came up. This gives you both, on OpenAI’s terms. It also creates a new channel to staff and measure, with a paid placement option attached, and the agent you put there will speak for your brand in a setting you don’t control.
Now What: If you sell to consumers or small businesses, claim your profile and check that your product, pricing, and availability data are accurate, starting with what ChatGPT already says about you. Before launching a Business Agent, decide who owns its answers, test it on your hardest customer questions (returns, pricing exceptions, complaints), and set up a way to review what it tells people. Track mentions and clicks for a quarter before you put ad budget behind it.
Running Agents in Production
Four stories this week are about the system around the model: where agents run, how their output gets tested, what they remember between sessions, and who sells the tooling to oversee them. That system decides whether agents hold up in production.
Anthropic Offers a Self-Hosted Version of Claude Managed Agents, in Private Beta
What: Anthropic announced on November 12 that Claude Managed Agents, its service for building and running long-running agents, can now be self-hosted, in private beta. In the self-hosted version, the customer runs every part of the service in its own cloud or on-premise: the agent harness, the sandboxes where agent code executes, credentials, and session history, which is stored in the customer’s own database. Model requests go where the customer chooses, whether to Anthropic directly, to Amazon Bedrock, Google’s Agent Platform, or Microsoft Foundry, or through the customer’s own gateway. Anthropic describes four design principles that apply to every deployment option: the harness is separated from the sandbox, credentials stay out of the sandbox, network access can be limited to a list of allowed hosts, and every session leaves a full record that exports over OpenTelemetry. Access is through Anthropic account teams.
So What: Agents that fix builds, triage alerts, and work through bug backlogs need access to source code and internal systems, and many security teams won’t approve that when the control plane sits in a vendor’s cloud. Until now the alternative was building your own agent platform. Self-hosting gives regulated companies a third option, with the trade that you now operate the server, database, and workers yourself.
Now What: If a security review has blocked an agent deployment over where code, credentials, or session logs would live, bring this to your security team and ask your Anthropic account team about the beta. Anthropic’s four principles also work as a checklist for any agent platform you evaluate or build, whoever the vendor is: separation of harness and sandbox, credentials kept out of the sandbox, an allowlist for network access, and a complete exportable record of every session.
AWS Releases a Free Local Emulator of Its Cloud, Written Entirely by AI Agents
What: AWS released AWS Emulator on November 11, a free tool that simulates more than 200 AWS services on a developer’s own machine so that developers and coding agents can test code without a network connection or cloud charges. GeekWire reported that AWS engineers did not write or review the code. “The emulator was written entirely by autonomous agents. Our team hasn’t written or reviewed any of the code,” wrote Matt Goldberg, the AWS senior principal engineer behind the project. Agents work from AWS’s public API models in three teams: one writes tests and checks them against the real AWS cloud, one runs those tests against the emulator and files issues, and one fixes them, with every change passing an automated gate for coding standards and security. Goldberg wrote that hundreds of agents now fix thousands of issues a day. An IDC analyst told InfoWorld the emulator doesn’t yet simulate throttling, outages, or regional differences and doesn’t sandbox the code it runs. The agents’ list of known gaps is public.
So What: AWS shipped software its own team didn’t write or review, and its case for doing so rests on the setup around the agents. AWS vice president Swami Sivasubramanian put the engineer’s job as building “the specs, tests, and merge gates that decide when that work is correct.” The emulator itself also addresses a cost many teams already feel, since agents that test by running code against real cloud services add delay and spend to every loop.
Now What: Try the emulator for local test loops if your coding agents work against AWS, and keep a final pass against real AWS for the behaviors it doesn’t cover. For your own agent-written code, pick one class of change (test fixes, dependency updates) and write down what specification, tests, and automated gate would have to exist before it merges without human review.
Anthropic Research: Agents Remember Better When They Review Their Own Notes Offline
What: Anthropic published research on November 6 on how Claude builds and maintains memory across sessions. Anthropic reports that newer models save far less and make better use of it: playing Pokémon Red, Claude Sonnet 3.7 wrote 413 memory files and earned one badge, while Claude Fable 5 wrote 10 files and earned four. The post describes “dreaming,” a background process in which Claude checks its saved notes against session transcripts, corrects errors, and builds linked summaries the researchers call wikis. In one test on about 350,000 internal documents, answers using both search and an agent-built wiki met 61% of graded criteria, against 53% for search alone. The post also reports failures: one wiki grew larger than the documents it summarized, and when several agents wrote to one shared page they overwrote each other, and only 41 of 57 graded facts survived until the instructions were changed to give each contract its own page. Telling models to “record only failures” reduced the benefit memory gave them. Anthropic also notes that notes loaded into every session use up context and that longer instruction files reduce how closely the model follows them.
So What: Memory is what lets an agent get better at your work over time, and the research lists how it goes wrong: notes that were wrong when written, notes that go stale, and summaries that bloat. The fixes Anthropic describes cost compute, and someone has to decide what gets checked and how often. For the strongest models, a simple statement of what memory is for worked better than detailed rules about what to save.
Now What: If you run agents that keep notes or instruction files between sessions, schedule a regular review pass that checks those notes against what happened and prunes what’s stale. Keep the always-loaded layer short. When several agents write to shared knowledge, give each page one owner. Before investing in an agent-built knowledge base, test it on questions that need several documents to answer, which is where Anthropic saw the gain.
LangChain Raises $300 Million at a $3 Billion Valuation to Build Agent Oversight Tools
What: The Wall Street Journal reported on November 12 that LangChain, which makes a platform for building AI applications and agents, raised $300 million at a $3 billion valuation, more than double its valuation in October of last year. CapitalG, Alphabet’s independent growth fund, led the round. LangChain plans to use the money to expand LangSmith, its enterprise product for monitoring, evaluating, and deploying agents. “The agent’s behavior is the hard part to control,” chief executive Harrison Chase told the Journal. Workday’s chief AI officer said the company uses LangChain’s tools to build the agents in its platform. LangChain has 435 employees, including 47 engineers who work directly with customers, and plans to reach 850 by the end of next year. The Journal also cited PitchBook data showing AI venture funding fell 43.4% in the third quarter from the second, to $94.3 billion.
So What: Investors are funding the control layer for agents (monitoring, evaluation, and deployment) as a business separate from the models. The pitch from LangChain and its lead investor is independence: the model vendors and cloud providers each offer agent tooling tied to their own models and infrastructure. If you expect to use more than one model vendor, where your agent logs, evaluations, and deployment controls live is a decision with lock-in consequences of its own.
Now What: Before choosing how to monitor and evaluate agents, list what you need to see for every agent run (inputs, tool calls, outputs, cost, and who approved what) and check whether each option gives you that across all the models you use. Ask whether your evaluation data and traces can be exported in a standard format. Decide this before your second or third agent goes live, when switching gets expensive.
Blank Metal is an AI consulting and engineering firm. We help organizations move from AI experiments to production systems. Learn more