Weekly AI News — Google changed three models and Search in five days, OpenAI handed its agents a wallet, and United put "the customer's agent" outside the door

Three big announcements this week. Google opened Gemini 4 Carbon to all paying users on November 10, released a Decisions API on the 11th that "chooses without writing an answer," and shipped Gemini 4 Hydrogen on the 12th while putting it into Search for every user worldwide the same day. OpenAI released ChatGPT Wallet on the 13th so an agent can finish the payment, with a person approving each purchase by biometric check and the card number hidden behind a token. On the other side, United Airlines issued a press release on the 11th blocking bookings by "unauthorized, third-party personal AI agents," and from the 12th Meta's Muse calls US businesses on behalf of paid subscribers. Anthropic's Claude Fable 5.2 and Mythos 5.2 (the 9th) were covered separately in NO.44, so we add only a short note, and we close with Ramp's October AI Index as "In brief."

This article was written with AI (Claude) and reviewed by an editor. Our company is a paying customer of Anthropic, the maker of Claude, and of OpenAI; we also test Google models, and we run apps, commerce and AI channels ourselves. Item 1 is a Google announcement and item 2 an OpenAI announcement, both competitors of Anthropic; item 3 is Anthropic's own announcement; and the Ramp data in "In brief" shows Anthropic ahead of OpenAI. So we add no evaluation of any side and only carry the numbers. Company-reported figures are marked "as reported by the company," and anything from a single source is marked with that source. This is not an investment opinion about any company and not material for investment decisions.

It was a busy week. NO.44 (Fable 5.2's two doors), NO.45 (United, Businesses in ChatGPT, Carbon public) and NO.46 (Hydrogen's price) each took one item per day, so this weekly note places those three back into the shape of the week. Over five days Google split its model slots into three and added a fourth slot that "only chooses." OpenAI let an agent finish the payment but left the final approval in human hands. The same week an airline blocked the customer's agent, and Meta's agent started calling businesses. The "door" question (NO.02, 34, 41) came down to the "wallet" this week.

1. Google — Carbon public, the Decisions API, Hydrogen, and Search, in five days (facts)

In date order. On November 10, Gemini 4 Carbon opened to all paying developers and enterprises in Google AI Studio, the Gemini API and Gemini for business, and reaches Google AI Ultra subscribers through Antigravity, the Gemini app and AI Mode in Search (as reported by the company). The price is $2 per million input tokens and $10 per million output tokens until November 30 (the same as Argon), then $4 and $20. The announcement says "Argon remains the ideal choice for the majority of software engineering and knowledge work."

On November 11, Google released the Gemini Decisions API (model name Gemini Decision 1) (as reported by the company). It does not generate an answer token by token; it only scores options set in advance. Three question types — yes/no (the probability that a condition holds), choice (pick one of 2 to 255 options and return probabilities), and rubric (per-level probabilities and a weighted score) — with up to 1,000 questions per request, inputs up to 256K tokens, and text, images, audio, video and PDFs mixed in one input. Every answer carries a confidence from 0 to 1, and the post says to use it to hand ambiguous cases to a person. The price is $0.10 per million input tokens and $0 for output, and the shared input is billed once per request. The company's figures — 150–250 ms per request (median 50–100 ms of model compute), nearly the same latency for 1 and 100 questions, the lowest calibration error on its own benchmark suite, and roughly 5x cheaper than Gemini 3.5 Flash-Lite (Interactions API) — are all as reported by the company. Its examples are model-tier routing in Gemini Notebook, picking one emoji out of 200-odd, and a safety layer in Google Search that screens adversarial behavior in MCP tools. It can be tried on AI Studio's no-cost tier.

On November 12, Gemini 4 Hydrogen arrived (NO.46 covered its price and positioning — $1.50 and $7.50 per million tokens, "Carbon and Argon for depth, Hydrogen for everyday execution," four thinking levels). The same day a post by Robby Stein, VP of Google Search, said "Hydrogen is now available globally in Google Search for all users" and called it "the biggest model upgrade to Search yet." What changed: understanding of long, multi-part or typo-filled questions; a query fan-out that splits one question into sub-questions and surfaces a wider range of sites; more images, comparison tables and custom tools; and better answers in every supported AI Mode language. There are no figures in that post.

For the independent side, we opened the Artificial Analysis leaderboard ourselves on November 14. NO.46 said "Hydrogen is not on the leaderboard yet"; now it is — Hydrogen xhigh 54, high 51, medium 41, low 27 (the same numbers three outlets had carried). The top is Claude Fable 5.2 (max) 65, then Fable 5.2 xhigh and high 62, Gemini 4 Carbon (high) 60, Claude Opus 5.5 (max) 59, Carbon (medium) 58, Opus 5.5 xhigh and GPT-6 Astra (max) 57, Fable 5.2 medium 56, and Astra xhigh and high, GPT-6.1 Sol (max) and Claude Sonnet 5.5 (max) at 55. The leaderboard page shows no index version number or date. An Anthropic model sits at the top, so here too we add no evaluation.

2. OpenAI — ChatGPT Wallet: the agent finishes the payment, but "a person approves every purchase" (facts)

OpenAI released ChatGPT Wallet on Friday, November 13. The announcement did not open on our side, so we confirmed it through two outlets, TechCrunch and The Decoder. The structure: the user registers a card once; ChatGPT's agents ("dots") find and pick the product; just before payment the user sees the items, the total and the selected card; and the user approves each purchase with Face ID or a fingerprint. Payment runs on payment tokens from American Express, Mastercard and Visa (stand-in numbers for the real card), so neither the agent nor the merchant sees the real card number, and OpenAI says card rewards and the right to dispute stay as they are (the company's words). The agent passes the merchant only what the order needs, such as the shipping address and contact details.

The scope is narrow. It opened first to a limited number of ChatGPT Pro subscribers in the US, aged 18 or over, with a US-issued Visa, Mastercard or Amex card, and the company called it a limited launch "to test and improve." What it previewed: buying without per-purchase approval within a set spending limit, purchase category and time period, pre-approval of specific purchases, and more supported retailers. For merchants it published a "Commerce Partner Playbook"; that the playbook covers firewall adjustments, product data and payment providers (Stripe, Adyen, Riskified), and that OpenAI dropped its earlier Instant Checkout as "not flexible enough" and rebuilt, are from The Decoder alone (the commerce page we opened at developers.openai.com shows only the Agentic Commerce Protocol — product feeds, promotions, a measurement pixel, a conversions API — and no playbook text). Neither outlet mentions fees. TechCrunch added that Amazon blocks agents and steers users to its own Alexa (citing Business Insider), and that DoorDash and Uber have reported few agent orders (NO.35, 40).

The other side of the same week. On November 11, United Airlines published a press release in its newsroom titled "United Restricts Access for Third-Party AI Agents" — the statement NO.45 described as "carried by three outlets, original not found," which we opened directly this week at the location TechCrunch gave. The content matches the quotes: sharing account credentials with personal, third-party AI assistants can expose sensitive information and lead to unintended purchases, account changes or use of MileagePlus miles; automation can breach the human acknowledgment that FAA and Department of Transportation rules require (such as baggage declarations at check-in); so United will prohibit use by "unauthorized, third-party personal AI agents or assistants" directly on its digital platforms and indirectly through travel agencies and aggregators; it is updating booking, ticketing and distribution policies, adding guidance to its Digital Trust Center, and will keep evaluating how personal AI assistants interact with its services. There is no effective date. In the same week sit CBS's analyst quote ("flight searches cost airlines money," and Kayak, Hopper and Google Flights stay as they are) and two headlines on Skift's list (the CEO of hotel group MCR saying he wants to ban AI bots from booking, and Citi writing that AI companies should pay when their agents book the wrong travel — bodies not opened).

And from November 12, Meta's Muse calls US businesses on behalf of paid US subscribers ($20 a month for 500 million tokens a week, $100 for 3 billion) — the cable company, the DMV, a restaurant with no online reservations. What was limited to a few applicants in September is now open to all paying users. But the primary source is an X post by Meta's chief AI officer Alexandr Wang (TechCrunch and MakeUseOf cite the same post; there is no post in Meta's newsroom), and nowhere is it written how Muse tells the person on the other end that it is an AI. MakeUseOf added that earlier reports described human contractors handling some calls in Meta's tests, and that Meta has not said whether people are involved in the current version (one source).

3. Anthropic — Claude Fable 5.2 and Mythos 5.2, and the old price still in the docs (facts, conflict of interest)

The November 9 announcement was covered in NO.44 — one model released through two doors with different safeguard thickness (Fable for everyone, Mythos through a verification program), the $8 and $40 price, the losing cells in the company's table. Two facts to add this week. First, we opened Anthropic's model documentation on November 14 and Fable 5.1 and Mythos 5.1 are still listed at $10 and $50 — the figure NO.44 flagged as "the 20% cut is our own calculation" is now confirmed inside the same document. The docs still recommend "Opus 5.5 ($4 and $20) as the default for most workloads," and describe Fable 5.2 as for "demanding reasoning and long-horizon agentic work, or when Opus 5.5 at higher effort still falls short." Second, the leaderboard in item 1 above — the independent score went from "none yet" (November 10) to "65." The announcing company makes the AI that wrote this note, so we add no evaluation and write nothing about its listing.

In brief — Ramp's October AI Index: adoption up, token prices down (facts, conflict of interest)

Ramp published its October AI Index on November 11 (Ramp's data page, opened directly; NO.43 covered the September edition). Among US businesses (more than 85,000 Ramp customers, by payment records), the share paying for AI services was 56.8% (+1.0 point, slightly faster than September's +0.66). By vendor: Anthropic 44.8% (+1.1), OpenAI 40.6% (flat), SpaceXAI 6.3% (+2.2). Total AI spend rose 1.65% from the month before to a new peak, but while the top 1% of firms spend a median $10,200 per employee per month (+9%), the median firm spends $16.32 per employee (+$0.93), and Ramp's chief economist repeated that "1% of businesses drive 80% of the revenues to OpenAI and Anthropic." In the new department-level data, engineering is 72% and customer success and support went from 0.8% in June to 13.3%. The average price per million tokens fell 21% in a month, and token volume grows faster than spend. Token-spend growth was 2.1% for OpenAI and 0.5% for Anthropic (The Decoder's summary, also in Ramp's post). We disclose that Anthropic is ahead on some figures, and the sample caveat (skewed to tech and finance, "adoption is likely overstated") is Ramp's own.

Reading the three together (interpretation)

First, Google's slots are four, not three. After Carbon (hard tasks), Argon (most work) and Hydrogen (everyday), there is a Decisions API that "only chooses, without writing an answer." Many things a small team runs every day — is there an item to avoid on this label, which desk does this inquiry go to, what grade is this review — are judgments that need no generated answer. Carving those out into a call with $0 output tokens changes the "cost per task" arithmetic (NO.46) again (interpretation). But the latency, calibration and "5x cheaper" are all company figures, and the cost of a wrong judgment stays on our side.

Second, the last step of the wallet is still a human hand. NO.45 said "neither seller nor platform has yet handed over payment, irreversible actions or human confirmation." This week OpenAI handed over payment but kept per-purchase approval with a person, hid the card number behind a token, and left automatic purchases within a limit as a "preview." The picture of the week is that the reasons United blocked (credential exposure, unintended purchases, human confirmation) and what OpenAI designed (no card number exposure, per-purchase approval) address the same items (interpretation). The side closing the door and the side opening it are looking at the same list. And Muse's calls leave one line on that list blank — "does it say it is an AI?"

Third, the number of companies using AI grows slowly, token prices fall fast, and usage shifts toward support. Put Ramp's three numbers (+1.0 point, −21%, support 13.3%) next to the three model announcements and the message is that spend is decided less by a new model's price list than by "which tier for which job" (interpretation). The test table in NO.17 and NO.46 is the table that does that job. Ramp's numbers are Ramp customers' numbers, not ours.

What we would do — checks to run this weekend (practice)

Add a "judgment-only call" column to the test table. Of the four-model test in NO.46 (label judgment, app copy, Lia script draft), take only the label judgment, run the same 20 inputs through the Decisions API in yes/no form, and write the confidence and "our reference answer" next to each result. AI Studio's no-cost tier is enough. The comparison columns are three — accuracy, the share of low-confidence cases, cost per task — and Safe List's principle of "never saying safe" (NO.03) stays: the judgment may only say "no item to avoid was seen."
Rewrite the draft rule for incoming agents against the three sentences in United's original. NO.45 proposed a "why" column (credentials, irreversible steps, human confirmation); now that the original exists, we organize the draft paragraph for our commerce terms around those three items and put it to the CEO for a decision. We add one line in the opposite direction — when an AI calls our customer service number, "ask whether it is an AI, and record the answer with the call." The same sentence whether it is Muse or Gemini.
Check the price-change dates in the rate table and recount with "our token receipts." Put Carbon on December 1 (2/10 → 4/20), Flash on January 1, 2027 (NO.46, VentureBeat alone) and Fable 5.1 → 5.2 (10/50 → 8/40, confirmed in the docs) at the top of the table, and since Ramp's "−21%" is an average and not our price, write last month's actual price per million tokens from our API bill next to it. ChatGPT Wallet is less something we will use than something our store will receive, so we note in one line each what our commerce already has and lacks among product feeds (Agentic Commerce Protocol) and payment-provider integration.

What we did not pick this week

Reports on the change of leadership at Nvidia (many on TechCrunch) touch corporate governance, personal allegations and the share price, so we record only the headline. A Fed governor's warning about an "AI investment correction" and reports on Anthropic's listing are investment material and were not used. TechCrunch's report (November 12) that a breach at Fireworks AI exposed Notion customers' chat transcripts is a security incident, headline only — we will follow up on NO.32 once an official notice appears. Alibaba's Qwen4-Max-Preview (official docs: a preview limited to November 11–25, 1M context, OpenAI- and Anthropic-compatible interfaces, no per-token price or benchmarks) waits for the final price after the preview ends. OpenAI's Businesses in ChatGPT (November 10) was covered in NO.45 and its announcement still does not open. ByteDance's Muse-like agent (Bloomberg alone), Nectar's creator-discovery agent (TechCrunch alone), Google DeepMind's AlphaGenome 2 and Gemini Robotics 2.5, and Anthropic's Claude for Manufacturing and Google Meet integration (conflict of interest) stay as headlines. Stanford HAI's post on the AI100 report (November 10) is a follow-up candidate for the NO.13 timeline. Learn:US No.38 (five free GitHub repos to cut subscription costs) is a follow-up candidate for NO.20 "check before you install."

What we could not confirm

OpenAI's ChatGPT Wallet announcement and the Commerce Partner Playbook text did not open. "Instant Checkout was dropped" and the Stripe/Adyen/Riskified integrations are from The Decoder alone, and fees, merchant costs and the list of supported stores appear in neither outlet. The size of the "limited number of Pro subscribers" was not disclosed.
Muse's calling feature has Alexandr Wang's X post as its primary source and we did not open the original post (TechCrunch and MakeUseOf cite it). As of November 14 there is no such post in Meta's newsroom. How it discloses being an AI, and call limits, appear nowhere.
The Decisions API's latency, calibration error, "5x cheaper than Flash-Lite" and the effects in its examples (40% lower latency in Notebook) are all company-reported. We have seen no independent evaluation yet.
The hotel (MCR) and Citi items on Skift's list were seen as headlines only; we did not open the bodies. United's press release has no effective date.
The Artificial Analysis leaderboard page shows no index version or evaluation date. We did not open Google's pricing page this time either, so the Flash increase in 2027 remains the single-source (VentureBeat) note from NO.46.
Ramp's 2.1% (OpenAI) and 0.5% (Anthropic) token-spend growth appear with the same numbers in Ramp's post and The Decoder, and the "69.5% share of spend for top-tier models" we saw directly in Ramp's post.

What we read

We opened and checked sources of different kinds. Company announcements and official documents, research and evaluation bodies and data, press coverage, wikis, and our earlier articles are listed separately.

Company announcements and official documents:

Research and evaluation bodies, data:

Press coverage:

Wikis:

Our earlier articles: