Connect with us

NEWS

Google Opens Gemini 4 Carbon at Argon’s $2 Rate

Google opened Gemini 4 Carbon at Argon’s $2 intro rate as a long-horizon coordinator, while Fairwind still gates the cyber build.

Published

on

Google opened Gemini 4 Carbon to paying developers and Google AI Ultra subscribers on November 10 at $2 per million input tokens, matching Gemini 4 Argon. Koray Kavukcuoglu, Google DeepMind’s chief AI architect, called it the firm’s strongest model for cyber defense and long-running agent work across coding, research and knowledge jobs.

Argon stays the everyday pick. Carbon is the heavier one, built to spend more tokens checking its own work, and the one Google is seating as the lead agent when a job splits into a team.

Google Opened Carbon to Developers on November 10

Kavukcuoglu’s November 10 post said Google is making Gemini 4 Carbon available to developers, businesses, scientists and consumers. The same post still tells teams to keep Argon on most software and knowledge work, and to bring in Carbon for multi-step debugging, open-ended research and jobs that “test the limits of a model’s intelligence.”

Carbon is based on Argon. It takes text, images, audio and video, and it ships with a combined 1M token context window for input and output. Thinking depth is tunable at low, medium and high. The knowledge cutoff is March 2026. The API id is gemini-4-carbon.

The public door opened after a gated start. Google announced Carbon on October 28 and first gave it to a subset of cyber defenders, then said it would gather feedback on guardrails before a wider release.

THE CARBON ROLL OUT

  1. September 2, 2026: Google launches the Fairwind Program so vetted governments and partners can use its most advanced cyber-defense models.
  2. October 28, 2026: Kavukcuoglu announces Gemini 4 Carbon and starts Fairwind access, with a public release described only as coming “as soon as possible.”
  3. November 10, 2026: Carbon reaches the paid Gemini API, Gemini for business, Google Antigravity, the Gemini app and AI Mode in Search for Ultra subscribers.
  4. November 30, 2026: Last day of the intro token rate.
  5. December 1, 2026: Input and output prices double to $4 and $20 per million tokens.

Google AI Studio’s own note on launch day stated the intro window in those dates, and said cached input stays 95% off the input rate.

The Same $2 Rate as Argon, on Purpose

The list price is the Argon list price. Through November 30, Carbon is $2 per million input tokens and $10 per million output tokens. From December 1 those figures become $4 and $20. Cached input is 95% off, so $0.10 per million during the intro window and $0.20 after.

TOKEN PRICES FOR CARBON

Token type Through November 30 From December 1
Input, per 1M $2 $4
Output, per 1M $10 $20
Cached input $0.10 (95% off) $0.20 (95% off)

Same sticker, different meter. Google says Carbon uses more tokens than Argon to explore, reason and run extra check loops. A hard job at $2 can still cost more than an Argon job at $2, because the heavier model keeps thinking. That is the bill teams will actually see, not the headline rate.

Google’s own launch copy already warns people off treating Carbon as a default. The model is for the hard slice of the workload. Hands-on tests on November 10 and 11 showed the same pattern: Carbon can take longer wall-clock time and more tokens than Argon or Claude Fable 5.2 on a one-shot 3D build, even when the result looks better. The cheap-looking rate is a per-token rate. It is not a promise that a long agent run will be cheap.

The practical setup, then, is routing. Put Carbon on plan, design and review. Put a lighter model on the file edits. That is also how Google is wiring Antigravity.

Put Carbon on Architecture and Auditing

Google is previewing /swarm in Google Antigravity on the same day as the API opening. Type /swarm and one prompt becomes a family of agents: a coordinator, specialist workers in isolated git worktrees so edits do not collide, and verifiers that run tests and challenge diffs. They post to a shared, append-only message board. An idle agent parks at zero token cost and wakes when another agent needs it.

The product is not “one smarter chat.” It is a small factory. Carbon is the expensive chair. Argon, and even Gemini 3.8 Flash, are the hands.

HOW A SWARM IS MEANT TO SPLIT

  • Coordinator: Carbon takes the brief, splits work and keeps the plan in bounds.
  • Workers: Faster models edit in separate worktrees so parallel changes do not overwrite each other.
  • Verifiers: Separate agents run tests and argue with the diffs before anything is treated as done.

Google DeepMind engineer Philipp Schmid wrote that moving from Argon is a model-id change to gemini-4-carbon, and that he uses medium thinking_level for almost everything. For bigger multi-file jobs, he said, “You can put Carbon on architecture and code review while 3.8 Flash works on parallel edits.”

Inside Google, engineers dogfooded /swarm for five weeks on multi-package migrations, cross-subsystem root-cause hunts and architectural redesigns. The Antigravity team’s warning is blunt: even with zero-token idle waits, parallel multi-agent runs chew quota much faster than a normal chat. The Swarm UI tries to make that visible, with a live Board, a Members roster and a Graph timeline you can scrub.

Fairwind Still Holds the Cyber Keys

The public Carbon and “Carbon for cyber” are not the same door. Specialized find-exploit-patch tools stay inside the Fairwind Program for trusted defenders. Google launched that program on September 2 for governments, critical-infrastructure operators and other vetted partners. Defenders still apply. The November 10 opening did not put the cyber build on the public API.

On Google’s October 2026 table, Carbon for cyber leads the defense loop Google cares about: 88.2% on SEC-bench Pro for finding flaws, 54.8% on ExploitGym for proving they can be exploited, and 71.0% on CWE-bench for patching. Argon sits at 83.4%, 42.0% and 68.0% on those three. GPT-6 Astra is 85.4% and 42.4% on the first two. Those cyber rows, the model card notes, reflect Fairwind access.

Early use inside Google is already more specific than a leaderboard. In Project Fortify, Google’s scan-triage-remediate system, Carbon hit 100% recall on scanning and 95% precision on triage on an internal benchmark, and more than a 25% relative lift over Argon’s recall on open-source repos. Internal tests also show more than a 2x lift over Argon at writing secure code. The Chrome team said Carbon produced high-quality patches more than 10X better than another lab’s large frontier model.

Outside Google, the Cloud Vulnerability Research team used Carbon to prove two bugs in a popular open-source project were exploitable, in the same severity band as the PHP2Shell flaws, and to build a proof of concept that got the bugs reclassified as critical. As part of Wiz’s Scan for Good work, the model found 0-day vulnerabilities at two major SaaS providers in under two hours. Google says the Wiz team then worked with those firms to patch, covering potential impact on tens of thousands of companies.

In a controlled evaluation of kernel driver vulnerability research, every model we tested found the same core memory-corruption bugs. Carbon for cyber went further, it understood the driver’s constraints, separated the findings that hold up from those that don’t, and uncovered four additional verified vulnerabilities. That is the difference between a finding and an actionable result.

Chris Corde, Chief Product Officer, SentinelOne, Google DeepMind blog

Google also used the October 28 announcement to publish safety numbers that sit next to those cyber scores. Carbon recorded a 0.0% attack success rate on Gray Swan’s Indirect Prompt Injection test, against 0.7% for Argon. It cut successful cheating on Terminal-Bench 4.0 to 0.0%, with cheat attempts at 0.4%. High-severity misalignment alerts in live internal coding dropped by almost 80%. Overreach fell 26% and under-service fell 29%.

Where Carbon Leads, and Where It Does Not

Google’s line is “strongest model.” The tables are narrower. Carbon leads several agent and knowledge benches at a lower cost per task than the next outside model, Google says more than 2X lower, and it is not first on every coding or computer-use test.

Cursor’s own CursorBench 4.0 coding-agent tasks, drawn from real multi-file sessions, put Carbon High at 60.4% on November 10, just ahead of Claude Fable 5.2 Max at 60.0% and Claude Opus 5.5 Max at 57.8%. Gemini 4 Argon High is 52.9% on that board. Cursor did not publish Carbon’s cost per task. Fable 5.2 Max is listed at $32.79 a task and Opus 5.5 Max at $13.43, so Google’s “more than 2X lower” claim cannot be checked from Cursor’s cost column alone.

Perplexity’s WANDR research-agent run is the cleaner cost picture. Carbon scores 0.857 at $5.86 per task. Claude Fable 5.2 scores 0.851 at $14.51, about 2.5 times the spend for a slightly lower score. Zapier’s AutomationBench, as listed on Google’s model card, has Carbon at 53.9% and Argon at 51.3%, with Opus 5.5 at 42.5%.

CARBON AGAINST ARGON AND RIVALS

Benchmark Gemini 4 Carbon Gemini 4 Argon Best listed rival
AA Index v5.0 59 50 58, Claude Opus 5.5
CursorBench 4.0 High 60.4% 52.9% 60.0%, Claude Fable 5.2 Max
AutomationBench 53.9% 51.3% 42.5%, Claude Opus 5.5
SEC-bench Pro 88.2% 83.4% 85.4%, GPT-6 Astra
FrontierSWE 2 61.2% 55.0% 65.5%, GPT-6 Astra
OSWorld-2.1, offline 81.5% 74.7% 87.2%, Claude Opus 5.5

GPT-6 Astra still leads FrontierSWE 2 at 65.5% against Carbon’s 61.2%. Claude Opus 5.5 still leads OSWorld-2.1 at 87.2% against 81.5%. Carbon does take SWE-sweep at 9.6% versus Argon’s 4.3%, Terminal-Bench 4.0 at 65.2%, GraphWalks (256k to 1M) at 96.2% versus Argon’s 84.2% and Astra’s 71.8%, and the Artificial Analysis Intelligence Index v5.0 at 59, one point over Opus 5.5. A new Multimodal Knowledge Agent bench, built with Artificial Analysis, has Carbon at No. 2 and more than 6X lower average cost per task than the top model.

WebDev Arena is the other honest row. Carbon jumped 89 Elo to 1768 and landed fourth, not first. Google is selling a specialist, not a clean sweep.

Finance Desks and Overnight Builds

The customers Google chose to quote are not chatbot users. They are people who need files, charts and patches that survive a review.

Gemini 4 Carbon is the strongest model we’ve evaluated for deep financial research, building cleaner, better-structured Excel models and producing sharper PowerPoint decks. It is a real step forward for diligence workflows.

Shanks Salwan, Lead Applied AI Engineer, Hebbia, Google DeepMind blog

In pre-release tests, Carbon matched the best model Hebbia tried on its end-to-end financial agent eval, and ranked first among 23 frontier configurations on Balyasny Asset Management’s General Financial Intelligence benchmark, covering legal questions, tool use, coding, data work and reading charts. Audrey Liu, a senior AI engineer at Balyasny, said Carbon stood out on complex context and structured deliverables, “especially charts, code, and tables.” Glean’s Thai Tran said it led every frontier model the firm tested on task completion inside workplace data.

The coding specimen is overnight work, not a demo snippet. Kit Langton at OpenCode said his team handed Carbon a requirements list, closed the day, and found a new Rust UI library plus a rewritten desktop app with working demos in the morning. He now prefers it as a planning model for the hardest jobs. Baki Celtik at HubX said Carbon resolved 94% of 16 real bugs in a 14,000-test codebase with zero regressions. Andrew Carr at Cartwheel ran it for six days on a cloth simulation and said it improved speed and fidelity with little steering.

Swarm demos follow the same long-job shape. One team turned seven city-council planning PDFs, 206 engineering drawing views, into a 906K-triangle Three.js model of Google’s Platform 37 office across 10 overnight loops. Another rebuilt a playable c. 1780 Stein-Schule fortepiano from CAD and X-ray plates. Google’s launch video walks through an interactive 3D songbird vocal organ that synthesizes six species live.

Who Can Use Gemini 4 Carbon Now?

Paying API customers can call it today. Google AI Ultra subscribers are getting it in Antigravity, the Gemini app and AI Mode in Search. Carbon for cyber is still a Fairwind application. Scientists can apply for six months on Antigravity’s highest usage tier at no charge. There is no free public API tier in Google’s launch note.

THE DOORS THAT OPENED ON NOVEMBER 10

  • Paid API: Gemini API in Google AI Studio and Gemini for business, model id gemini-4-carbon, at the Argon intro rate through November 30.
  • Google AI Ultra: Antigravity app, CLI and IDE extensions, plus the Gemini app and AI Mode in Search, with /swarm in preview in the app and CLI.
  • Android Studio: Listed by Google AI on launch day as a developer surface alongside AI Studio.
  • Fairwind: Separate apply-and-wait track for Carbon for cyber, unchanged by the public API opening.
  • Labs and funds: Scientists apply for six months free on Antigravity’s top tier; Cursor, Perplexity and Databricks all added the model on November 10.

Databricks said Carbon scored 66.7% on OfficeQA Pro v2 in its enterprise-document tests, against 57.8% for Argon. Perplexity put it on Pro and Max plans, including persistent Teammates and one-off agent sessions in Perplexity Computer. Cursor added it at the same $2 / $10 intro rate, rising on December 1.

Consumer access is the tighter door. Carbon in the Gemini app and in Search is an Ultra feature. The API is open to any paying developer. Those are different products at the same model name, and the cyber tools are a third product still behind a background check.

From December 1 the token meter reads $4 and $20. Until that date, the coordinator and the workhorse cost the same per token, and the extra spend shows up only if you let Carbon do work Argon could have finished.

Harry is the editor and lead writer of CUMBERNAULD MEDIA, which he runs as an independent publication after a decade in journalism spent moving from reporting to editing. His habit is to open the document before the summary of it. A company result is read from the filing rather than the press release, a court or regulatory decision from the judgment itself, a scientific finding from the paper and its methods section rather than the headline claim, and a sporting sanction from the governing body's own ruling. That approach shapes coverage across news, business and technology as much as science, sports and entertainment, and it carries into the lifestyle, travel, auto and gaming pages, where product specifications are checked against the manufacturer's sheet and, where possible, against Harry's own testing. Every number is checked before publication, and where a source's figures are disputed the story says so. Corrections follow a public policy and are marked on the page. Readers anywhere in the world who write in get a reply from him, and the address is support@cumbernauld-media.com.

Click to comment

Leave a Reply

Your email address will not be published. Required fields are marked *

Trending