Google releases Gemini 4 Carbon, its strongest model, at 2x lower cost per task
Google released Gemini 4 Carbon on Nov 10 to all paying developers and enterprises, calling it its strongest model — aimed at long-horizon agentic work and cybersecurity — and saying it ranks #1 on CursorBench 4.0, Perplexity's WANDR and Zapier's AutomationBench at more than 2x lower cost per task than the next-best external model.
📊 Scorecard: our Bullish call on $GOOGL is scored from $287.38, the last close before it ran. Its first result comes after the next close.
The News
Google released Gemini 4 Carbon on November 10 to all paying developers and enterprises, through its API — the programming interface developers use to build on a model — and Gemini for business, as well as to Google AI Ultra subscribers in the Gemini app, Antigravity and AI Mode. Google calls Carbon its strongest model, aimed at long-horizon agentic work — tasks where an AI agent works on its own over a long stretch, such as running a multi-step project — and cybersecurity.
The benchmark claims are striking, with a caveat that matters: they are Google's own framing of third-party tests. Google says Carbon ranks number one on CursorBench 4.0, a coding benchmark from the Cursor team; Perplexity's WANDR, a test of research work; and Zapier's AutomationBench, a test of business workflow automation — at more than two times lower cost per task than the next-best external model. A benchmark is a standard test used to compare models, and the rankings here are Google's presentation of results from tests run by other companies, not an independent league table.
Pricing is aggressive. Introductory pricing is $2 per million input tokens and $10 per million output tokens — a token is roughly a word fragment, the unit AI companies charge by — the same as Gemini 4 Argon, rising to $4 per million input tokens and $20 per million output tokens on December 1. Google is also previewing a feature called /swarm in Antigravity, its developer workspace, which turns one prompt into a coordinated team of agents working together.
Top of three benchmarks at more than 2x lower cost per task — Google is fighting the model race on price, not just capability.
Why It Matters
This matters because the AI model race is now fought on cost per task, not just capability. A model can be the smartest in the world, but if it costs too much to run, developers will not use it for real work. Google's claim — top of three benchmarks at more than two times lower cost per task than the next-best external model — is aimed squarely at that calculation. If it holds up in independent testing, it gives developers a reason to switch, and switching costs in AI are real: once a team builds around a model, it tends to stay.
It also matters because of where Carbon is aimed. Long-horizon agentic work and cybersecurity are the two areas where AI companies believe the next wave of value sits: agents that can carry a project for hours instead of answering one question, and security tools that can find flaws faster than attackers can. Releasing Carbon the same week as Gemini Robotics 2.5 — which is powered by the Gemini family — shows Google pushing the same model line into coding, research, security and now physical robots at once.
The pricing schedule is worth reading carefully. The introductory $2 and $10 rates double to $4 and $20 on December 1, so the cheap window is three weeks long. That is a classic land-grab: get developers building on Carbon now, at a price that undercuts rivals, and bet that by December they are too invested to leave. The risk for Google is that the benchmark claims, which are its own framing of third-party tests, do not survive independent scrutiny — rivals publish their own numbers within days in this race, and Anthropic's own benchmarks, published the same week, already show its Fable 5.2 beating Carbon on some tests.
Analyst's Take
Our call is Bullish. Google's strongest model, top of three third-party benchmarks by its own account, at more than two times lower cost per task than the next-best external model, with introductory pricing that matches its previous flagship — that is a strong hand. The December 1 price rise is the tell: Google wants developers locked in before the cheap window closes, and it is confident enough to set a deadline.
What to watch: independent benchmark results in the coming days, and whether developers actually switch at the introductory price. Alphabet closed Tuesday at $287.38, down 0.27% — the market has not yet reacted to Carbon, which means the reaction, if there is one, comes when usage numbers, not launch headlines, start to show.
Are you bullish or bearish on this story?
By The Numbers
FAQ
What is Gemini 4 Carbon?
It is Google's newest flagship AI model, released Nov 10 to paying developers and enterprises. Google calls it its strongest model, aimed at long-horizon agentic work — tasks where an AI works on its own over a long stretch — and cybersecurity.
What does 'more than 2x lower cost per task' mean?
Google says Carbon costs less than half as much per completed task as the next-best rival model, while ranking first on three third-party benchmarks. That is Google's own framing of those tests, not an independent ranking.
How much does it cost to use?
Introductory pricing is $2 per million input tokens and $10 per million output tokens — a token is roughly a word fragment, the unit AI companies charge by. From December 1 the price rises to $4 per million input tokens and $20 per million output tokens.
What is /swarm?
It is a feature Google is previewing in Antigravity, its developer workspace. It turns one prompt into a coordinated team of AI agents that work together on a task, rather than a single agent working alone.
Who can use Carbon now?
All paying developers and enterprises through the API and Gemini for business, plus Google AI Ultra subscribers in the Gemini app, Antigravity and AI Mode.
👁
Source: Google