Three dollars per million input tokens. Fifteen for output.
That is the list price for Gemini 4 Neon, released November 18.
From December 1, flagship Carbon costs $4 and $20. Against top rivals, Google (NASDAQ: GOOGL) says Neon's cost per task is 50% to 80% lower; "half the price" refers to that, not to the sheet.
Third place is not the story
On the Artificial Analysis intelligence index at the xhigh setting, Neon scores 61 and ranks third. Ahead: Claude Fable 5.2 at 65, GPT-6.2 Astra at 62.
Third is not news. The schedule is. Four product changes in seven weeks: Carbon on October 28, Hydrogen on November 12, Neon on November 18, with 3.1 Pro retired in between. Rankings cannot sit still at that cadence. Google is not waiting for an all-round champion. It is filling price points.
61 is a composite: it flattens reasoning, coding, math and instruction-following into one number, and a long-tail failure disappears into the average. 61 says Neon has no obvious weakness, not that it has none.
Elsewhere: Vals Index 70.9%, second. Agent Arena Index 1336, second, and second on steerability. AA's coding agent index 58.6, first. Three seconds and a first describe the position better than the total: it holds second, and prices at the second tier.
Where the strength is
Terminal-Bench 4.1: 65.2%. FrontierCode 2.0 at the default high setting: 46.2%. Closer to daily work. Sub-scores are the hands; the composite is the physical.
One caution. The xhigh setting is each vendor's top thinking configuration, so a cross-vendor comparison needs the rival at its top tier too. Otherwise the four points between 61 and 65 may only be a difference in thinking budget. SWE-sweep scores 39.0% in the Antigravity environment and 12.8% under the mini-SWE harness: same model, 26 points apart. Ask which harness produced a number before which rank the model holds.
Google's own scores
The safety and cyber figures look excellent: Terminal-Bench 4.0 refusal rate 1.98%, cheating 0.0%; Gray Swan indirect prompt injection 0.0% attack success at the top thinking setting. Five firsts in cyber: AutoBaxBench hard tier 76%, SEC-bench Pro 90.5%, CWE-bench v2 70.7%, WebExploitBench v1.2 51.26%. Neon for cyber will open to trusted defenders through the Fairwind program.
All of these are Google's own evaluations, and third-party retesting is not out yet. In-house evaluation is not fake, it is a standard: you write the questions, build the environment, choose the baseline, and the numbers look good. The two kinds sit side by side like one report card. They are two different exams.
Expensive per token, cheap per task
The list price is higher than Carbon's: 3 against 2, 15 against 10, both 50%.
But on December 1 Carbon moves to 4 and 20, and Neon becomes 25% cheaper. Same model, two dates, two conclusions. The price sheet has no "recommended" label. The spread writes it.
Argon, the previous workhorse, also moves to 4 and 20, so renewals feel it first. Google's other number matters: at the default high setting, cost per task is 32% below Carbon. Price 50% higher per token, 32% lower per task, and only one explanation fits. Neon burns far fewer tokens for the same work. Back out the arithmetic and consumption per task is about 45% of Carbon's. That is my arithmetic, not Google's; it holds only if both do the same kind of task, and the task set is unpublished.
Divide the price gap by the score gap: four points for 25%, about six points per point. The conversion is sloppy, but the point holds. The flagship premium is not charged by score. It is charged by the cost of being wrong.
Long context is where the per-token price bites. Context and output both reach one million tokens; text, images, video, audio and PDF go in. A million input tokens at $3 is $3; a cache hit is $0.15, 5% of list. Read the same material ten times and the total is $4.35, not $30. The catch: caching needs a stable prefix. Change one character and the hit rate collapses, so the saving also dictates the method: fix the material first, then ask.
The rival inside the house
Neon sits in the middle: Carbon the hardest projects, Hydrogen the high-frequency work.
Two numbers squeeze Carbon. Neon is 32% cheaper per task and first on the coding agent index. Carbon keeps the hardest tasks; the question is how many there are. A mistake there is expensive to redo, so buyers care about first-pass success, not list price. Most requests do not need 65. They need 61 at a lower price, and 61 is already third.
This is designed self-cannibalization. In the Gemini app, Neon replaces Argon and 3.1 Pro retires, leaving three price tiers instead of five models. The flagship no longer exists to sell the most units. It sets a ceiling for the middle tier and proves capability before pushing it down. Carbon launched October 28; Neon followed three weeks later.
The real rival is not Fable 5.2. It is Carbon.
Taking Carbon's users needs no advertising. It needs two numbers ordered on a price sheet and four entry points: Gemini API, Gemini for business, Antigravity and the Gemini App.
Who still buys 65? Tasks where a mistake is expensive. For writing code, fixing bugs and reading documents, 58.6 on the coding index means more than a 65 composite. Procurement reads sub-scores; the boss reads the total bill.
My read: the anchor is moving from dollars per million tokens to dollars per finished task. The per-task measure has a soft spot: it depends on a task set Google defined. Change the set and 32% could become 12%, or 52%.
Open question: once 61 is good enough, how long can the 65 premium hold?
*Disclaimer: This article is based on publicly available information and is for reference only. It does not constitute investment advice.*
