FORSMILE
Issue #14Published November 22, 2026 / Covering Nov 16-22日本語で読む

This Week in AI, 9 Items — OpenAI and Google Update Their Flagship Models in the Same Week, Anthropic Completes Its IPO, and NVIDIA Reports $115.8B in Revenue

The week in one place

Three threads this week. The first is OpenAI and Google updating their flagship models in the same week. OpenAI's GPT-6.2 Astra costs $8 per million input tokens and $40 output, the same per-token price as last issue's Claude Fable 5.2. OpenAI says it beats Fable 5.2 by 5.5 points "at just 5% of the cost", which is a claim under conditions OpenAI chose. Google's Gemini 4 Neon is $3 and $15, and Google positions it between Carbon and Hydrogen by workload (Carbon is $2 and $10 as an introductory price until 30 November, then $4 and $20). In the same week Genie 4 reached eligible Google AI Ultra subscribers and, a day later, developers through the API. The second is money and compute moving. Anthropic closed its IPO at $83 a share (a follow-up to last issue's watching list), and also closed a concurrent private placement of about $26.0B with AMD, Broadcom, Google, Micron, Microsoft, NVIDIA and TSMC Partners. NVIDIA reported third-quarter revenue of $115.8B and left data center compute revenue from China out of its next-quarter outlook. The third is AI moving into design and rule-making. OpenAI published its chip-design results with AMD and a new benchmark for medicinal chemistry, and New York's attorney general created a unit to enforce the state's AI safety law that takes effect in January 2027.

Models & Products

OpenAI releases GPT-6.2 Astra and Sol: Astra is $8 input and $40 output, and OpenAI claims it beats Fable 5.2 by 5.5 points at 5% of the cost

OpenAI's top model and its everyday model were updated together. Both the scores and the costs use comparisons OpenAI chose.

On 17 November OpenAI introduced GPT-6.2 Astra (its frontier model for the hardest projects) and GPT-6.2 Sol (its efficient model for everyday work). In ChatGPT Work and Codex they are available to all Plus, Pro, Enterprise and Edu users (Enterprise and Edu once a workspace admin enables them), and in the API as `gpt-6.2-astra` and `gpt-6.2-sol`. API prices per million tokens: Astra is $8 input, $0.40 cached input and $40 output, which is 20% lower than GPT-6 Astra on input and output and 60% lower on cached input. Sol keeps GPT-6.1 Sol's $2 input, $0.10 cached input and $10 output. On Agents' Last Exam, which tests multistep professional work, OpenAI says Astra scores 5.2 points above GPT-6 Astra and 5.5 points above Claude Fable 5.2 at 5% of the cost (the Fable 5.2 results exclude nine tasks that used a fallback model). On Cognition's FrontierCode 2.0, Sol comes within 0.5 points of Astra at about one-fifth the cost (both at max).

So What

Astra's per-token input and output prices are the same as Fable 5.2's ($8 and $40) from last issue. With the same unit price, the "5% of the cost" difference may come from how many tokens are used or from the conditions, and I could not read the breakdown from the page. The scores are OpenAI's own tally, so whether you get the same gap on your work needs separate testing.

Models & Products

Google releases Gemini 4 Neon: $3 input and $15 output, a 1 million-token context, positioned between Carbon and Hydrogen

Google places Neon between Carbon and Hydrogen. Its unit price is above Carbon's introductory price and below Carbon's standard price, and Google says its cost per task is lower.

On 18 November Google introduced Gemini 4 Neon, for coding, analysis and enterprise work. Google positions it this way: Carbon takes your hardest long-horizon projects, Hydrogen handles fast everyday execution, and Neon balances depth and efficiency between them. It is priced at $3 per million input tokens ($0.15 cached) and $15 output, with a 1 million-token context window. Google says it scores 61 on the Artificial Analysis Intelligence Index at xhigh thinking (#3 overall) and ranks #2 on the Vals Index on the Pareto frontier. At the default high thinking level it has a 32% lower cost per task than Carbon on the Artificial Analysis measure. It is rolling out in the Gemini API (Google AI Studio), Gemini for business, Google Antigravity for paid AI Pro and Ultra subscribers, and the Gemini app for Pro and Ultra subscribers, and Neon for cyber goes to eligible Fairwind Program members over the next few days.

So What

Carbon is $2 and $10 as an introductory price until 30 November, then $4 and $20 (a footnote on Carbon's announcement). Neon's $3 and $15 sit above the introductory price and below the later one. "Lower cost per task than Carbon" is a claim on measures Google chose, so your own cost needs separate measuring. The cyber version goes only to limited defenders and is handled apart from the general API.

Models & Products

Google rolls out Genie 4 in Project Genie for eligible Ultra subscribers and, a day later, in the Gemini API: rewrite a world by typing as you move, now with sound

The world model announced as a research preview on 27 October reached eligible Ultra subscribers and developers this week.

Genie 4 is a world model that Google DeepMind announced as a research preview on 27 October. On 18 November Google began rolling it out in Project Genie for eligible Google AI Ultra subscribers. While you move around, you can type new prompts to change the weather, add a character or redirect the action, and in Google's words Genie 4 pairs the visuals with real-time native audio that adapts to the scene (ambient sound and dialogue) for the first time. Project Genie also gains an agent that builds a world through conversation and a gallery for sharing, and Google AI Pro subscribers can join a waitlist. On 19 November Google wrote that Genie 4 is available to developers through the Gemini API (via the Interactions API) and Google AI Studio, where you can generate 720p, 24fps worlds with synchronized native audio from a text prompt, a reference image or both. Outputs carry SynthID watermarks. Google treats Genie 3 and 4 as frontier research models and lists camera latency, intricate physical interactions, real-time audio precision and long-term scene permanence as challenges it is still working on.

So What

The vision (training environments for games and robots) and the limits Google admits today (such as long-term consistency) are in the same documents. API pricing is on a separate Google page that I did not check. Project Genie is limited to the "eligible" Ultra subscribers Google describes.

Funding & Corporate

Anthropic closes its IPO at $83 a share: about $93.5B in gross proceeds, plus a concurrent private placement of about $26.0B with seven companies

The price and listing date that last issue left on the watching list are now settled. The price came in above the $70 to $78 range expected earlier.

On 18 November Anthropic announced it priced its initial public offering at $83.00 per share, and shares began trading on the Nasdaq Global Select Market on 19 November under "ANT". In its 20 November closing announcement, the offering totaled 1,127,000,000 shares (including the full exercise of the underwriters' 147,000,000-share option), of which 1,103,351,354 were sold by Anthropic and 23,648,646 by selling stockholders. Gross proceeds were about $93.5B, of which about $91.6B went to Anthropic. It also closed a concurrent private placement of 313,253,007 Class N shares at the same $83 to AMD, Broadcom, Google, Micron, Microsoft, NVIDIA and TSMC Partners, for about $26.0B. The IPO plus the placement came to about $119.5B. When the roadshow launched on 10 November, the plan was 825,000,000 shares at an expected $70 to $78.

So What

The terms moved from the tentative roadshow plan (825,000,000 shares at an expected $70 to $78) to a completed offering (980,000,000 shares at $83 when priced, 1,127,000,000 after the option). The amounts are gross proceeds before expenses, and Anthropic's portion is separate from the selling stockholders'. I have not covered the trading price after listing because Anthropic's own announcements do not include it (and I did not verify it). The announcements do not mention any change to model prices or contracts.

Funding & Corporate

NVIDIA's third-quarter revenue is $115.8B, up 103% on a year ago; its $130.0B outlook for next quarter includes no Data Center compute revenue from China

Demand for AI compute continues. Data Center compute revenue from China is left out of the outlook.

On 17 November NVIDIA reported results for the third quarter of fiscal 2027, which ended 25 October 2026. Revenue was $115.8B, up 20% from the previous quarter and up 103% from a year ago. Data Center revenue was $107.6B, up 21% quarter on quarter and up 110% year on year. The GAAP gross margin was 74.8%. The outlook for the fourth quarter is revenue of $130.0B, plus or minus 3%, and NVIDIA says it is "not assuming any Data Center compute revenue from China". It says Vera Rubin NVL72 is the fastest product ramp in its history, with racks running at partners including CoreWeave, Google Cloud, Microsoft Azure, Oracle Cloud Infrastructure and Nebius. It also says it expanded Anthropic's access to NVIDIA Grace Blackwell and Vera Rubin capacity through cloud providers.

So What

The key point in the numbers is that the outlook assumes no Data Center compute revenue from China. NVIDIA is also a buyer of shares in Anthropic's concurrent private placement this week (item 4), and says it expanded Anthropic's access to its compute. A seller of AI compute also being a shareholder in a company that uses it is visible across the official documents.

Research

OpenAI publishes its chip-design results with AMD: up to 14% improvement on a flagship GPU design, and GPT-6.2 Astra tops its RISC-V tests

Circuits rewritten by the model have started to enter designs sent for manufacturing. The limits of the tests are in the same article.

On 18 November OpenAI published how its models improve chip design. With AMD it saw up to a 14% improvement in power, performance and area (PPA) over heavily optimized baselines on a flagship GPU IP, which OpenAI calls roughly a year of progress in under five months. AMD has already included some of the changes in a recent design sent for manufacturing and verified single-digit percentage PPA gains on key blocks of its upcoming flagship CPU. On OpenAI's own chip Jalapeño, model-proposed changes cut standard-cell area by about 13% in the vector unit and about 10% in the matrix unit, and these were silicon proven. In tests on two open-source RISC-V cores (9,881 scored runs), GPT-6.2 Astra had the highest average score on both at a budget of about 112,000 tokens. OpenAI notes that designing from a specification lets a model omit features of the original cores, and that a more representative test would require those features, a tighter area limit and a timing check after physical synthesis.

So What

The claim that some changes entered designs sent for manufacturing and the admission that the tests are easier than real chip design sit side by side in the same article. The improvement figures are those shown by OpenAI and AMD, with no third-party verification in the article. OpenAI says it is working with Synopsys to integrate with design tools (GPT-Synopsys).

Research

OpenAI releases MedChemBench for medicinal chemistry: 99 problems, GPT-6.2 Astra scores 84.5% and GPT-5.6 Sol 58.2%

A new yardstick for reading molecular structures, interpreting experiments and choosing what to test next. The score comparison is among OpenAI's own models.

On 20 November OpenAI introduced MedChemBench: 99 problems drawn from medicinal chemistry research, using 117 papers and patents and 761 distinct molecular structures. It grades four things: reconstructing molecular structures, evidence-based explanation, ranking and selecting compounds, and predicting numerical values and categories. At the xhigh reasoning setting, GPT-6.2 Astra earned 84.5% of the available credit, GPT-6 Astra 83.3%, GPT-6.1 Sol 82.0% and GPT-5.6 Sol 58.2%. GPT-6 Astra was strongest at reconstructing structures and explaining experimental evidence, and selecting compounds and predicting numerical properties proved harder. One example problem draws on inhibitors of PRMT5, a protein studied in cancer drug discovery.

So What

The only models with scores in the text are OpenAI's own. Comparisons with other companies' models cannot be read from this article. It is not a number that decides usefulness in a lab; it is a new yardstick for how far AI can handle molecular structures and experimental results.

Developer Tools

Claude Code 2.1.326 to 2.1.330: sign-in to Enterprise Admin Console organizations, easel for drawing charts, and auto mode on Bedrock and others

More admin-side settings and features for using it through cloud providers.

Between 16 and 20 November, Claude Code shipped 2.1.326 through 2.1.330. 2.1.330 lets you, after `claude auth login`, sign in to organizations managed in Anthropic's Enterprise Admin Console and apply that organization's settings and its cloud provider or gateway. It also adds easel, a built-in mod that draws charts on Pro, Max and, where available, Team and Enterprise on macOS and Linux (automatic charts can be turned off in `/config`). 2.1.328 adds auto mode for Claude Opus 4.6 and Sonnet 4.6 on Amazon Bedrock, Google Cloud's Agent Platform, Microsoft Foundry and the Claude apps gateway, and `"onFailure": "block"` for `mcp_tool` hooks, so a hook whose call fails blocks the action. 2.1.327 adds `claude mcp add-json <name> -`, which reads an MCP server's JSON from stdin so tokens can stay off the command line. 2.1.329 fixes `~/.claude.json` sometimes being reset to defaults when a network stall or file lock made it unreadable.

So What

The changes centre on people using Claude Code under a company's cloud or admin console. The auto mode entry names Opus 4.6 and Sonnet 4.6 specifically, so it does not mean it has spread to every latest model. Check each release's notes for your environment and the release dates.

Regulation & Safety

New York's attorney general creates the ASSET Unit to enforce AI laws, including the RAISE Act that takes effect in January 2027

A dedicated unit for enforcing AI safety laws now sits in a state attorney general's office.

On 17 November New York Attorney General Letitia James announced the AI Safety, Standards, Enforcement, and Transparency (ASSET) Unit. It will lead enforcement of the Responsible AI Safety Education (RAISE) Act, which takes effect on 1 January 2027. The announcement says the RAISE Act requires AI developers to publicly disclose their safety measures and swiftly report security and safety incidents. The unit will work on data privacy, digital scams, cybersecurity and online safety, and develop policy proposals. The announcement lists the laws the office can use: a data security law, the RAISE Act (which regulates large frontier AI model developers), a ban on unsafe chatbot features for children that takes effect in June 2027, and a law requiring AI chatbots to address suicidal ideation and to tell users they are not talking with a human.

So What

The announcement describes New York State laws and duties on AI developers; it does not describe any duty on people using AI from Japan. But with an enforcer now named for a law that asks frontier-model developers to disclose safety measures and report incidents, more developer disclosures may follow from January 2027 (the announcement does not say that they will).

Watching (no confirmed primary source yet)

  • Microsoft's Maia 200 announcement at Ignite (17 November) appeared only in third-party summaries; I have not opened Microsoft's own page, so it is not an item.
  • Reports that California's attorney general sued OpenAI, the US Commerce Department's interim rule on AI chip export licenses, and remarks by President Trump and President Xi came only from third-party roundups, and I have not opened a primary source. New York's announcement refers to "reports of AI agents at OpenAI hacking another AI company", but I have not opened the source of that.
  • Anthropic's share price after listing is not in Anthropic's own announcements, and I did not verify it, so I have not covered it.
  • OpenAI's "Inside the Codex sandbox" and "Managing frontier AI risk", and Anthropic's feature on Claude helping plan a drive for the Perseverance rover, were seen only as titles in listings; I have not opened the pages (and have not confirmed the feature's date).
  • Google says Gemini 4 Neon for cyber will roll out "over the next few days", and I did not check whether it has completed. I also did not check whether the score and cost claims for GPT-6.2 Astra and Sol have third-party verification.
  • I could not load official primary sources for xAI, Meta, Mistral, DeepSeek or Alibaba (Qwen) this time. That means unconfirmed, not that there were no new announcements.

"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