FORSMILE
Issue #13Published November 15, 2026 / Covering Nov 9-15日本語で読む

This Week in AI, 12 Items — Anthropic and Google Ship Top and Mid-Tier Models in the Same Week, as Anthropic's IPO Prep and Agent Shopping and Payments Take Shape

The week in one place

Three threads this week. The first is the top model and the price tiers below it moving in the same week. Anthropic's Claude Fable 5.2 is $8 per million input tokens and $40 output, Google's Gemini 4 Carbon is $2 and $10, and Gemini 4 Hydrogen is $1.50 and $7.50, all prices and evaluations stated by the companies themselves. Thinking Machines Lab published Inkling 2 Small weights under Apache 2.0 as promised (a follow-up to last issue's watching list). The second is agents moving closer to shopping and money. OpenAI introduced ChatGPT Wallet, in which an agent pays without holding the real card details, and for business added Businesses in ChatGPT and ChatGPT Enterprise through AWS Marketplace. The third is trust and verification. Anthropic started its IPO roadshow, NVIDIA donated OpenShell, which constrains what agents can do, to the CNCF, and Google DeepMind used AI agents to formally verify the actual code of Signal. A study by OpenAI and Upwork describes how freelance demand is shifting.

Models & Products

Anthropic releases Claude Fable 5.2 and Mythos 5.2: $8 input and $40 output per million tokens, FrontierCode 2.0 at 49.1% (Opus 5.5 at 45.5%)

Anthropic's top models move to 5.2, with Fable open to all users and Mythos to vetted organizations.

On 9 November Anthropic announced Claude Fable 5.2 and Claude Mythos 5.2. Fable and Mythos are the same underlying model; Fable has strong cybersecurity and biology safeguards and is available to all users. Organizations pursuing cybersecurity or biomedical work can apply for Mythos through the expanded verification programs. Pricing is $8 per million input tokens and $40 per million output tokens, and Anthropic says output is up to 30% faster than Fable 5.1 and Mythos 5.1 (an optional fast mode is up to 2.5 times faster, at $16 and $80). On the agentic-coding evaluation FrontierCode 2.0, it reports 49.1% at the Xhigh effort setting, against 45.5% for Opus 5.5 and 42.8% for GPT-6 Astra in the same table. It is available on all platforms, including Amazon Web Services, Google Cloud and Microsoft Azure.

So What

Every figure is Anthropic's own measurement; we could not confirm any third-party reproduction (it points to the system card for details). What matters in practice is the price and the guidance on when to use it: Anthropic itself advises starting with Opus 5.5 for most coding and knowledge work and moving up to Fable or Mythos as needed.

Models & Products

Google makes Gemini 4 Carbon generally available; API price is $2 input and $10 output per million tokens (introductory)

Google's top model reaches developers, enterprises and consumers.

On 10 November Google began offering Gemini 4 Carbon to developers, enterprises, scientists and consumers. It says it had released Carbon two weeks earlier to cybersecurity defenders through its Fairwind Program. The API is in Google AI Studio and Gemini for business for paying developers and enterprises, at $2 per million input tokens and $10 output (the same introductory price as Gemini 4 Argon). For Google AI Ultra subscribers it is rolling out in Antigravity, the Gemini app and AI Mode in Search. Google says Carbon ranks #1 on CursorBench 4.0, Perplexity's WANDR and Zapier's AutomationBench at more than 2X lower cost per task than the next best external model. Antigravity gets a preview of "/swarm", which turns one prompt into a coordinated team of agents.

So What

The rankings and cost figures are Google's own claims. With Anthropic releasing Fable 5.2 at $8 and $40 the same week, the top-model unit price now differs by a factor of four on both input and output. Cost per task also depends on tokens used, so unit prices alone do not compare.

Models & Products

Google releases Gemini 4 Hydrogen: flat $1.50 input and $7.50 output per million tokens across the 1M-token context, 2.5 times faster than Gemini 3.8 Flash

Gemini gets an "execution" model, available across all tiers.

On 12 November Google introduced Gemini 4 Hydrogen. It positions it as the model that runs coding and agentic workflows quickly, next to the frontier Carbon and Argon. Pricing is flat across the whole 1-million-token context: $1.50 input and $7.50 output. Against Gemini 3.8 Flash it says Hydrogen finishes tasks 2.5 times faster at its default setting and lifts Terminal-Bench 4.1 from 18.2% to 57%. On the Artificial Analysis Intelligence Index it scores 51 at its default high thinking level, above Argon's 50 (54 at the highest setting). It is available in all tiers of Google AI Studio, the Gemini API, Gemini for business, the Gemini app and Google Search, and for paid Google AI Pro and Ultra subscribers in Antigravity. Antigravity also gets "Orchestrator", which switches between models such as Carbon and Hydrogen within one session.

So What

These are figures quoted from Google's announcement, and we could not confirm third-party reproduction. The notable point is that putting a cheap, fast executor under a top model, a division of roles, was pushed to the front as a product in the same week by both Anthropic (Fable 5.2 as advisor or orchestrator) and Google (Orchestrator).

Models & Products

Thinking Machines Lab publishes Inkling 2 Small weights under Apache 2.0, a week after promising them: 594B total parameters on Hugging Face

The weights promised for "next week" in last issue's watching list were released as promised.

On 12 November Thinking Machines Lab wrote, in an update to its announcement page, that Inkling 2 Small's weights are now available for download on Hugging Face. Its 5 November announcement had said it would release them under Apache 2.0 after external safety testing and red-teaming. The Hugging Face repository (thinkingmachines/Inkling-2-Small) lists the license as apache-2.0, a total of 594,050,672,720 parameters, and is public without gating. The company also lists support in Unsloth, SGLang, vLLM and TokenSpeed, and API availability at NVIDIA, Fireworks, Together, Baseten, Nebius, Modal and Cloudflare.

So What

Last issue's watching list said we would check the license and the files once released. The license and the 594B scale are confirmed from the Hugging Face metadata. We have not confirmed that the weights perform as announced.

Developer Tools

Claude Code 2.1.321 to 2.1.325: Fable 5.2 becomes the default Fable, Code Mode arrives as an experiment, and an MCP credential leak and other issues are fixed

A model switch, an experimental feature and security fixes landed in one week.

2.1.321 (9 November) made Claude Fable 5.2 (claude-fable-5-2) the default Fable model (1M context, $8 input, $40 output and $0.20 cache reads per million tokens). The same release fixed MCP OAuth requests sending secrets such as refresh tokens to another host when a server redirected them, and a case where a read-only Bash command ran outside the sandbox without a permission prompt. 2.1.323 (12 November) added Code Mode, an experimental opt-in in which Claude writes JavaScript to call its tools. 2.1.325 (13 November) fixed a security issue in how zstd-compressed responses are decoded.

So What

If you use Claude Code, the practical point is that updating to 2.1.321 or later includes the secrets-handling and sandbox fixes. The release notes do not say how far the issues reached or whether they were exploited. Code Mode is experimental, and you turn it on yourself.

Models & Products

OpenAI announces ChatGPT Wallet: agents pay without holding the real card details, and each purchase needs a biometric or similar check

Agent shopping gets a payment method that does not hand over card details, for now only for US Pro users.

On 13 November OpenAI announced ChatGPT Wallet, built with American Express, Mastercard and Visa. An agent pays without access to the user's actual card number, and the user approves each purchase with a check such as Face ID or a fingerprint. Payment uses a token from the card network (a substitute for the card number, security code and expiration date), and OpenAI says card rewards and protections for eligible disputes are preserved. It starts as limited availability through dots for ChatGPT Pro subscribers in the US. OpenAI writes that over time it expects users to delegate purchases within set limits such as spending caps and purchase categories. For merchants, it says sites must make sure agents are not inadvertently blocked, and payment and risk systems must recognize legitimate agent purchases.

So What

Availability is limited to US Pro users, and the post does not say when users or merchants in Japan will be covered. For anyone running an online shop, the requirement not to block agent purchases automatically is now written out concretely in OpenAI's own document.

Market & Industry

OpenAI launches Businesses in ChatGPT in beta: companies register details and product feeds, see mentions and clicks, and in the US get a Business Agent

Companies can now manage their own information inside ChatGPT and have a conversation front door for customers.

On 10 November OpenAI started Businesses in ChatGPT in beta. In Business Hub, a company claims its organization (after verification) and updates its details and a self-service product feed with pricing and availability. Aggregated insights show mentions in ChatGPT and website clicks, plus discovery topics. OpenAI says it does not share conversations, and that insights appear only when counts meet certain thresholds. After claiming, a company can build and launch a Business Agent (an agent that answers questions and helps people explore offerings), starting with eligible ecommerce and online service businesses in the US. Shopify merchants can connect from Business Hub or the Shopify admin.

So What

For people running websites, it means there is now an entry point where a company itself can see its mentions and clicks in ChatGPT, not only search rankings, and correct its information (our reading of OpenAI's description). The Business Agent is limited to some US companies, and the announcement gives no figures on results.

Market & Industry

OpenAI puts ChatGPT Enterprise on AWS Marketplace; two days later AWS publishes a guide to using Codex through Amazon Bedrock

ChatGPT can now be bought on an existing AWS contract and bill.

On 9 November OpenAI announced that ChatGPT Enterprise is available through AWS Marketplace. Using the existing AWS purchasing and billing relationship, eligible spend can count toward AWS commitments. OpenAI says customers get the same models and features as buying directly, with new features released at the same time. Eligible customers can also access ChatGPT for Healthcare and ChatGPT for Financial Services through the Marketplace. In an 11 November technical post, AWS described Codex using Amazon Bedrock for inference with supported OpenAI models, and a small least-privilege pilot to start (the examples use GPT-5.6 Sol; it asks you to confirm that Sol, Terra or Luna is available in your AWS Region). Codex keeps running on the developer workstation.

So What

For companies that buy through AWS, two official routes now exist that use the existing AWS purchasing and billing path: buying ChatGPT Enterprise through Marketplace, and using Codex with inference through Bedrock. We have not confirmed pricing or availability by region, including Japan.

Funding & Corporate

Anthropic launches its IPO roadshow: 825,000,000 shares, expected price $70 to $78, Nasdaq listing applied for under "ANT"

After the 26 October S-1 filing, the terms of the offering are now concrete. They are not final.

On 10 November Anthropic announced that it launched the roadshow for its proposed initial public offering. The offering is 825,000,000 shares of Class A common stock (801,351,354 from Anthropic and 23,648,646 from certain existing stockholders), and the underwriters are expected to get a 30-day option to buy up to 123,750,000 more. The expected initial public offering price is between $70.00 and $78.00 per share, and Anthropic has applied to list on the Nasdaq Global Select Market under the symbol "ANT". Anthropic will not receive proceeds from shares sold by the selling stockholders. A registration statement has been filed but has not yet become effective.

So What

The price range, exchange and symbol are "expected" and "applied for"; the listing and final price are not settled (the announcement itself says shares cannot be sold, nor offers to buy accepted, before the registration statement becomes effective). This continues the 26 October S-1 filing covered in last issue's watching list.

Research

OpenAI and Upwork study freelance demand: in 43 of 63 work categories, activities common in ChatGPT lost share on Upwork; 20 showed the reverse

A study suggesting not simple replacement by AI, but changes in how work is bundled and in budgets.

On 10 November OpenAI published a study with Upwork. In 43 of the 63 work categories, activities more common in ChatGPT lost share in Upwork postings, and the other 20 showed the reverse. In AI-related postings, the mix of tasks became 45% more diverse between early 2023 and late 2025 (from about 15 equally common tasks to about 21), with more postings combining tasks that rarely appeared together before. Posted hourly rates were similar, but AI-related projects had larger fixed-price budgets and more hours of work (compared within the same type of work and the same posting month). OpenAI states that the findings do not point to a simple substitution story.

So What

The study rests on limited data, Upwork postings and ChatGPT usage patterns, and excludes work outside Upwork. For people who win contract work, it is material on estimate granularity: AI-related projects have larger budgets and different task bundles. No causation is shown (it also says ChatGPT use alone does not explain changes in freelance demand).

Developer Tools

NVIDIA donates OpenShell, a runtime that constrains what agents can do, to the CNCF; announced at KubeCon, and it also contributes to Google's Agent Substrate

A runtime that fixes what an agent may do by policy has moved outside a single company.

On 10 November at KubeCon + CloudNativeCon, NVIDIA announced that it is contributing NVIDIA OpenShell, an open source runtime for securing and governing autonomous agents, to the Cloud Native Computing Foundation (CNCF). As a CNCF Sandbox project it will be governed by the organizations building it into their products. OpenShell is described as applying deterministic controls so teams can enforce policies as agents use tools, access inference services and execute workloads, with connectable compute, inference providers, credential storage and monitoring tools. NVIDIA also says it is contributing to Agent Substrate, a Google-led open source framework for long-running agent workloads that has been accepted into the CNCF Sandbox.

So What

The trend of deciding what an agent may not do through policy in the execution layer, not through the model's output, has been placed in a vendor-neutral body (our reading of NVIDIA's announcement). The announcement gives no results on how much protection OpenShell actually provides.

Research

Google DeepMind formally verifies the actual code of the Signal protocol with AlphaProof Code: Gemini Carbon agents wrote over six million lines of Lean, and humans re-reviewed about 1%

DeepMind says AI mechanically proved the security of the implementation code, not only the protocol design.

On 10 November Google DeepMind presented AlphaProof Code. Against specifications that human experts formalized from Signal's security goals, agents powered by Gemini Carbon wrote proofs in over six million lines of Lean and mechanically proved that the actual code of the Signal protocol meets the specified security properties, it says. The agents audited the proofs and narrowed the areas needing a human second look to about 1% of the codebase. It describes the work as taking four weeks, for something that historically takes expert teams years. As a check, when two real, previously fixed bugs were reintroduced into an isolated version, the system produced a machine-checked proof that each buggy version was insecure. It says it is open-sourcing the formal verification artifacts.

So What

The framing as a "first" and a "grand challenge" is Google DeepMind's own wording. The scope confirmed is Signal's core cryptographic code against the specifications experts set; it does not mean the whole app is free of flaws. For someone covering security, the point worth noting is that a way to machine-verify implementation code has now been shown on a real, large example.

Watching (no confirmed primary source yet)

  • Gemini Decisions API (11 November), AlphaGenome 2 (11 November), Gemini Robotics 2.5 (10 November) and OpenAI's Control Room (9 November): we saw only the titles and dates on the official announcements and did not open the bodies, so they are not items.
  • We could not load the official feeds of xAI and Meta this time, and for DeepSeek and Alibaba (Qwen) our search returned only third-party roundups and we found no primary sources. That means we have not confirmed, not that there are no new announcements.
  • Anthropic's listing has no settled offering price or listing date. When they are fixed, we will open the primary sources (Anthropic's announcement or SEC filings) and cover it.
  • OpenAI's Images 2.6 and the expansion of GPT-6 Cyber to all customers, from last issue's watching list, were not re-checked this time.

"Primary" links go to the announcing party's own publication (company blog, press release, official docs). "Reporting" links go to news coverage or third-party analysis. Figures and dates are as verified on the publication date.

← All Weekly AI News issues