Google opens its strongest Gemini model at $2 per million tokens ahead of Anthropic's IPO pricing Nov. 18
Google's strongest model ships at the previous model's introductory price, a day after Anthropic cut its flagship 20 percent ahead of an IPO banks may push past $1.5 trillion. The contested ground is what a frontier token costs, and Google just named its price.
Vincent Jiangthree-minute read
A frontier model at last year's price
Google DeepMind opened Gemini 4 Carbon, its strongest model, to all developers, enterprises and consumers on November 10, after two weeks of cyber-only access through its Fairwind Program 1. The price is $2 per million input tokens and $10 per million output, the introductory rate of the older Gemini 4 Argon, doubling to $4 and $20 when the offer expires on November 30 1.
Google reports Carbon first on CursorBench 4.0, Perplexity's$30.00B — Perplexity, private, latest valuation $30.00B WANDR and Zapier's AutomationBench, at less than half the cost per task of the next best external model; those are the company's figures 1. It lands the November 8 phase-transition claims as a priced product, routed through Antigravity, Google's+2.69% — Google, up 2.69 percent today agentic coding platform, plus the Gemini app and AI Mode in Search, where revenue is subscription and cloud attach rather than token margin 1.
The price war Anthropic's IPO walks into
On November 9 Anthropic$350B — Anthropic, private, latest valuation $350B cut Claude Fable 5.2 twenty percent to $8 per million input and $40 output, while Meta−2.63% — Meta, down 2.63 percent today released Muse Ember, which runs benchmark tasks at about 14 cents against 41 cents for Claude Haiku 5.5 47. Anthropic prices its IPO on November 18 at $70 to $78 a share, worth up to $1.405 trillion fully diluted, and its banks are weighing a price above the range that could push the valuation past $1.5 trillion 56.
On the new 200-task Multimodal Knowledge Agent benchmark, built with Artificial Analysis, an independent benchmarking firm, Claude Fable 5.2 scored 56.56 percent and Carbon 54.34, a gap the technical report calls statistically indistinguishable 23. Cost is where they separate. Artificial Analysis priced Fable 5.2 at $31.97 per task and Carbon at $6.29; Google's report has Fable burning 40.5 million tokens a task against Carbon's 7 million 2.
Carbon's advantage is per-task cost, not accuracy
Data
| Value | |
|---|---|
| Claude Fable 5.2 | $31.97 |
| Gemini 4 Carbon | $6.29 |
| Claude Haiku 5.5 | $0.41 |
| Meta Muse Ember | $0.14 |
That record sits on a benchmark Google co-designed, with Carbon credited as an AI co-author 2. Google also says Carbon uses more tokens than Argon to reason deeper, so a held price means Alphabet absorbs the compute difference through November 1.
Carbon's introductory input price undercuts every rival's list price
The $2 intro rate undercuts every rival; the December 1 list price matches Opus 5.5
Data
| List input price per 1M tokens | |
|---|---|
| Gemini 4 Carbon (intro) | $2 |
| Gemini 4 Carbon (list, 1 Dec) | $4 |
| Claude Opus 5.5 | $4 |
| Claude Fable 5.2 | $8 |
| GPT-6 Astra | $10 |
Vulnerability hunting at a flat rate
Carbon ships with a security business attached. Google says Wiz, the security company, used it through its Scan for Good program to surface zero-day flaws at two major SaaS providers in under two hours, and that its internal Project Fortify hit 100 percent scanning recall and 95 percent triage precision, company figures throughout 1.
Anthropic moves the same direction: Mythos 5.2, its cybersecurity model restricted to vetted organizations, rediscovered 89.6 percent of 1,829 known vulnerabilities, and Claude Enterprise customers reach that flaw-finding through Claude Security 8. Frontier capability now arrives weekly from two labs; what separates winners from losers is who can still charge a margin for it. Google monetizes Carbon through subscriptions and cloud usage, Anthropic through the token price it just cut.
Two sides of a falling per-task price
Alphabet sits on one side of that split, converting frontier capability into distribution and cloud attach; pure model sellers and security-testing vendors face a falling per-task price. One caution travels with the bear case for the latter: a claim that a model demonstration halved software supply-chain company JFrog+5.44% — JFrog, up 5.44 percent today could not be verified in available reporting, and its November quarter beat estimates, revenue up 30.9 percent to $179.2 million 9.
The next readings are the November 18 IPO price and Carbon's list price on December 1 51.
Get Model by email, free
The AI labs, their models, and the business behind them. In our daily email, every weekday.
