The one-screen version: three of the biggest labs moved this week, and not one of them shipped a bigger model. Google released Gemini 4 Hydrogen (Nov 12), a fast, cheaper coding model priced at $1.50/$7.50 per million tokens — and bundled it with an Orchestrator that automatically routes work between models by difficulty, a vendor telling you to stop picking one model. OpenAI launched ChatGPT Wallet, letting its agents pay with tokenized cards and a biometric check on every purchase — agentic commerce with real card networks behind it, if narrowly gated. And Anthropic prices its IPO this week, expected to trade as 'ANT' around November 19. Each move changes the interface around the models, not the models. Here's what each asks of you.

1. Gemini 4 Hydrogen: the vendor just shipped routing#

Start with the one you can act on today. Google released Gemini 4 Hydrogen on November 12 — a speed-focused model built for coding and agent work that replaces 3.8 Flash and sits below the heavier Carbon and Argon tiers. It's pitched as the "daily driver for agentic workflows": debugging, repository-level edits, web dev, multi-step agent loops. API pricing is $1.50 per million input tokens and $7.50 per million output (reasoning tokens included), and Google claims it runs about 2.5x faster than 3.8 Flash at default reasoning. Independent early tests reported by partners put its Artificial Analysis Intelligence Index around 51 and a sizable jump on Terminal-Bench — figures that are Google's own or early-partner numbers, so verify on your workload before you quote them.

The headline for a founder isn't the benchmark. It's that inside Antigravity, Google's coding tool, Hydrogen ships paired with an Orchestrator that automatically routes each task to the right-sized model by difficulty. When the model vendor itself builds a router and sells it as the feature, that's the market telling you the era of pinning one model name in your config is over.

What it changes for you: put a cheap-fast tier under your agent loop and an escalation rule above it. Hydrogen is a strong candidate for the high-volume, latency-sensitive 90% of calls; reserve a heavier model — Gemini Carbon, Claude Opus or Fable — for the hard 10%. We've walked through building exactly this split in cut your LLM bill with task-based routing and a three-tier Claude router. The news just gave you one more cheap, fast default to drop into the bottom tier.

2. ChatGPT Wallet: the agent became a payer#

The quiet platform story of the week: OpenAI launched ChatGPT Wallet, which lets its agents — branded "dots" — complete purchases for a user without the agent ever seeing the real card number. Cards from Amex, Mastercard and Visa are tokenized, and every purchase requires a security step, a biometric approval. At launch it's narrow: ChatGPT Pro subscribers, 18 and over, in the US, with eligible US-issued credit cards (no debit). It's OpenAI's second run at agent-driven shopping after it wound down an Instant Checkout feature earlier this year.

Don't read this as "should I trust an agent with my card." Read it as a channel. A platform with hundreds of millions of users just wired real card networks to its agents and published guidance for businesses that want to accept those payments. That puts a concrete question on your roadmap: when an agent is doing the buying, is your product reachable, legible, and checkout-able by software instead of a human?

What it changes for you: probably nothing to build on yet — it's Pro-only and US-only, and agentic commerce is, in OpenAI's own words, early. But the habit it should start is real. Make sure an agent can find your product (structured data), understand it (clean descriptions and prices, not images of text), and complete a purchase (a checkout flow that doesn't dead-end a non-human visitor). The rails — ChatGPT Wallet, the agent-payment protocols, Visa's and Mastercard's agent programs — are converging, and the stores that are buyable by an agent get the orders the others never see.

3. Anthropic's IPO: today it's just the measuring stick#

The capital-markets event lands midweek. Anthropic is expected to price its IPO on November 18 and trade on Nasdaq as 'ANT' around November 19, at $70–$78 a share, an order book reportedly six times oversubscribed, for roughly $90B in total proceeds including a $26B private placement. We ran the full story — the lawsuit, the valuation, the week-long march to the bell — in yesterday's edition, so here it's the backdrop, not the lede.

Two things outlast the first-day pop, both of which we've made the case for before. It mints the first public comparable for a pure-play frontier-AI company, so your next raise gets benchmarked against ANT whether the comp fits or not. And its S-1 is a cost forecast — the filing's roughly $895B in compute commitments is the structural reason the model was never the moat and the reason your token bill isn't heading to zero.

The throughline: in five days, three labs changed the interface around their models rather than the models themselves — routing (Google), payment (OpenAI), a public price (Anthropic). Your unit of work moves up a level: from "which model do I call" to "how do I wire routing, payments, and a vendor I can underwrite." The discipline that survives all three is the one that's survived every month of this column — keep each layer swappable, so a new model, a new rail, or a new price is a config edit, not a rewrite.