The one-screen version: Google priced its new frontier model, Gemini 4 Carbon, at an introductory $2 per million input / $10 per million output tokens — exactly half of Claude Opus 5.5's $4/$20 — at rough parity with Opus 5.5 on Google's own intelligence index. The catch is the calendar: that rate expires November 30 and doubles to $4/$20 on December 1, landing right on top of Claude. So "half price" is a six-week promotion, not a new floor. And while the labs race the model price toward zero, the fresh venture money this week went somewhere else entirely — into the layer that governs the models. Here's what each half asks of you.
1. The price war is now three labs deep — read the ladder, not the headline#
Put this week's number in the line it belongs to. The frontier is now a price ladder at near-identical capability: GPT-6.1 Terra at $1/$5, Gemini 4 Carbon at $2/$10 (intro), Claude Opus 5.5 at $4/$20. A 4x spread on price, a rounding error on capability — Google's own published table places Carbon a hair above Opus 5.5 on a composite index, though independent benchmarks are mixed and that table is vendor-reported, so don't file it as fact.
The headline writes itself — "frontier intelligence just got cheaper" — and it's true, with an asterisk the size of the quarter. Carbon's $2/$10 is an introductory rate that, per multiple trackers, ends November 30; on December 1 it becomes $4/$20, exactly Opus 5.5's price. The "half" was a promotion. (And confirm you can actually buy Carbon tokens for your account and region first — reporting says Google opened it to paying developers in early November, but some trackers still flagged limited API access in the same window.)
What it changes for you: test now, price later. Run your evals on Carbon this week while it's cheap and migrate whatever passes — but build your own unit economics against the December 1 rate, not the promo. A cost floor that expires in six weeks isn't a cost floor; it's a trial. We said it when the deflation story broke: never sign a long token commitment into a market falling this fast, and never hand a customer a discount underwritten by a promo you don't control.
2. The model is the commodity now — stop shopping it like a moat#
Sit with the shape of that ladder for a second, because it's the whole strategy lesson. When three different labs will sell you near-parity frontier capability at a 4x price spread, and the spread is set by whoever's running a promotion this month, the model has finished becoming a commodity in public. The only rational way to buy a commodity is to keep switching to the cheapest acceptable supplier — which means the single most valuable piece of engineering you own is the swappable inference layer that lets you move labs without a rewrite.
This is the P&L finally saying out loud what we argued months ago: the model was never the moat. If your product is a thin markup on a token call, Carbon's price war is not good news for you — it's a preview of your own margin being competed to zero, because the margin inside the model is a melting ice cube and every lab is holding a hair dryer.
What it changes for you: treat model choice as procurement, not architecture. Abstract the call, re-run the eval at every renewal, take the cheapest tier that passes, and spend the engineering you save on the part a price war can't touch.
3. Follow the money — it left the model and funded the agent layer#
Here's the tell that makes the whole briefing cohere. In the same seven days three labs cut model prices, the fresh venture capital didn't chase a model — it funded the layer above the model. Bay came out of stealth on November 10 with about $50M (a $37M Series A led by Battery Ventures plus a $13M seed) to do one thing: vet and approve an AI agent's actions before they execute. A day later, Guava raised a $12.5M seed led by Acrew Capital for a developer platform to build voice agents.
Neither is a model company. One sells governance, the other sells tooling — and both got funded in the exact week the thing underneath them got cheaper. That's not a coincidence; it's the market repricing where value lives. The model is deflating, so capital is paying up for the layer that orchestrates, secures and specializes it. We flagged the opening move when agent security first got bought last week; Bay's raise says the category now has a funded floor.
What it changes for you: this is a map, so use it. If you're deciding what to build or where to put your next two engineers, the money is telling you: not on the token, on the layer that governs it. Agent orchestration, action-level permissions and audit, voice and vertical tooling, the proprietary data and workflow lock-in a customer can't get by calling the API themselves — that's the ground the price war can't flood.
The throughline: the cheapest way to read this week is as two arrows pointing opposite directions. The model price is falling toward zero in a three-way public race, and the venture money is climbing up — out of the model and into the governance and tooling layer stacked on top of it. Build on the falling thing and you're fighting gravity for a margin three labs are giving away. Build on the rising thing and you're where the capital already decided the value is. Pick your layer before December 1 picks it for you.
