NEWS
Google Opens Gemini 4 Carbon, Then Steers Users to Argon
Google opened Gemini 4 Carbon to paying developers and Ultra users, while saying Argon is still the better pick for most coding work.
Google opened Gemini 4 Carbon to paying developers on November 10, then told them Gemini 4 Argon is still the better pick for most work. Carbon is the stronger model. It is also slower, hungrier with tokens, and easier to overbuy.
Koray Kavukcuoglu, SVP of Google DeepMind and Google’s chief AI architect, wrote that Argon remains the ideal choice for the majority of software engineering and knowledge work. Carbon is for jobs that test a model’s limits, such as long debugging and open-ended scientific research, and it uses more tokens than Argon to do that extra thinking.
Google Tells Paying Users to Stick With Argon
The company is not hiding the split. In the Gemini app, Argon sits as the pick for code and analysis. Carbon sits as the pick for the most complex work. Both share a 1 million token context window, the same thinking_level controls, and the same API shape, so swapping them is a model-string change from gemini-4-argon to gemini-4-carbon.
Google’s own launch post on X put the hierarchy in one line.
If Argon is great for complex tasks, Carbon is the model you bring in for your hardest problems.
Google, official account, on X
Gemini 4 Carbon is now available.
It’s our strongest model, built to dig deeper, handle long-horizon tasks, and multi-step reasoning.
If Argon is great for complex tasks, Carbon is the model you bring in for your hardest problems. pic.twitter.com/4NVqHaMOPT
— Google (@Google) November 10, 2026
That line is a sales pitch and a warning. Teams that point every coding job at Carbon will pay for a long-horizon brain on work Argon is meant to finish faster. Two days after the paid opening, some Vertex routes were already returning capacity errors, and Carbon responses on other paid paths crawled. The model Google called its strongest is the one that is hardest to keep on the meter.
What Gemini 4 Carbon Costs on the API
Paying developers get Carbon in Google AI Studio and Android Studio at the same launch rates as Argon: $2 per million input tokens and $10 per million output tokens, with cached input at 95% off. After the introductory period expires on November 30, those rates become $4 and $20 on December 1.
GEMINI 4 TOKEN RATES
| Token type | Through Nov. 30 | From Dec. 1 |
|---|---|---|
| Input, per 1M | $2 | $4 |
| Cached input, per 1M | $0.10 | $0.20 |
| Output, per 1M | $10 | $20 |
The matching sticker price is the trap. Carbon is built to explore, reason, and run extra verify loops, so the same task can burn more tokens even before December 1. Artificial Analysis, scoring both models at the later $4/$20 list, put Carbon at $7.76 per Intelligence Index task and Argon at $7.68. The gap is small on that suite because Carbon also uses fewer output tokens than Claude Opus 5.5 (123k against 173k, 29% less). It is not small if you send Carbon work Argon would have closed in fewer loops.
THE CARBON ROLLOUT
- October 28, 2026: Google announces Carbon and opens it to a subset of cyber defenders through Fairwind.
- November 10, 2026: Carbon joins the paid Gemini API and begins rolling out to Google AI Ultra in the Gemini app, AI Mode in Search, and Antigravity.
- November 30, 2026: The $2 / $10 introductory rates expire.
- December 1, 2026: Input and output prices double to $4 and $20 per million tokens.
Google also ranks Carbon first on CursorBench 4.0, Perplexity’s WANDR, and Zapier’s AutomationBench at more than 2X lower cost per task than the next best outside model. That claim is about those three workflow tests, not the broader index, where GPT-6 Astra still undercuts Carbon at $6.25 a task.
A Sharper Guessing Habit on Harder Questions
When Carbon first appeared, Artificial Analysis placed it at 59 on the Intelligence Index v5, first among the models on that board. Argon scored 50. Claude Opus 5.5 scored 58. GPT-6 Astra scored 57. Carbon led four of the nine tests in that suite (HLE-Diamond, Terminal-Bench-Science, AA-Omniscience, and AutomationBench-AA) and was not leading on everyday coding and knowledge work.
INTELLIGENCE INDEX v5 AT LAUNCH
| Model | Index | AA-Omniscience | Hallucination | Cost per task |
|---|---|---|---|---|
| Gemini 4 Carbon (high) | 59 | 54 | 11% | $7.76 |
| Claude Opus 5.5 (max) | 58 | 48 | 50% | $10.04 |
| GPT-6 Astra (max) | 57 | – | 50% | $6.25 |
| Gemini 4 Argon (high) | 50 | 46 | 4% | $7.68 |
The knowledge test is the trade. Carbon’s accuracy on AA-Omniscience rose 10 points over Argon, to 58% from 48%, and its net score of 54 beat Opus at 48. Its hallucination rate at high reasoning rose from Argon’s 4% to 11%. That is still the second-lowest figure on the board, far under the 50% posted by Opus 5.5 and GPT-6 Astra. Carbon at a medium setting matches Argon’s 4%. The version Google is selling as its strongest brain is the one that answers more, abstains less, and invents more when it is wrong.
On the separate Coding Agent Index, Carbon in the Antigravity CLI scored 59, three points behind Claude Code with Opus 5.5 at 62. Google’s own Multimodal Knowledge Agent test, built with Artificial Analysis, has Carbon at No. 2 and more than 6X lower in average cost per task than the top model. The lead is real on hard, long jobs. It is not a reason to retire Argon.
Fairwind’s Locked Copy of Carbon
The public model is not the cyber model. Kavukcuoglu’s November 10 post says Gemini 4 Carbon for cyber, reached through Fairwind, leads on finding, exploiting, and patching flaws: 88.2% on SEC-bench Pro, 54.8% on ExploitGym, and 71.0% on Collinear’s CWE-bench. Those scores were run without guardrails. The public SKU is the one with the brakes on.
Google’s October 28 safety write-up said Carbon blocked every evaluated attack on Gray Swan’s Indirect Prompt Injection benchmark, a 0.0% success rate against Argon’s 0.7%. A later safety report on the public build says it blocked 99% of individual attack trajectories in Google’s own red-team tests. Those are different tests, and neither is the ungated Fairwind copy.
Inside Google, Project Fortify used Carbon for 100% recall on scanning and 95% precision on triage. Open-source repository tests showed an over 25% relative lift on Argon’s recall, surfacing logic flaws that had been missed. The Chrome team said Carbon was more than 10 times better than another lab’s large model at writing high-quality patches, and internal secure-code tests showed more than a 2x lift over Argon. Wiz, Google’s cloud security unit, used it in Scan for Good to find 0-day bugs in two large SaaS providers in under two hours.
Sundar Pichai, CEO of Google and Alphabet, said teams used Carbon with AlphaProof Code to produce over 6 million lines of machine-checked Lean proofs of the Signal protocol, compressing years of expert review into weeks, with about 1% of the codebase needing human eyes. That is the kind of job Carbon is for. It is also the kind of job most API customers will never run.
Fairwind itself launched on September 2 under Four Flynn, Google’s vice president for security and privacy. The program now lists more than 650 participating partners, with first place going to governments, national cyber authorities, operators of health, telecom, energy, and finance networks, and core technology platforms. Partners must keep access inside internal security, incident-response, or pen-test teams and use phishing-resistant multi-factor authentication. They may not share or resell the model.
One Prompt Spawns a Team of Agents
The other product hiding in the Carbon launch is not a chatbot. Google is previewing /swarm in Google Antigravity, which turns one prompt into a coordinator, parallel workers in isolated git worktrees, and adversarial verifiers that test every change. Agents talk on a shared, append-only board and park at zero token cost when idle. You can pin Carbon to architecture, coordination, cyber audit, and verification, and leave file edits to cheaper models.
Google’s demos are greenfield spectacles: an Eiffel Tower 3D show from old plates, a 906K-triangle rebuild of a Google office from city PDFs, a playable 1780 fortepiano from CAD and X-rays. On a live product repo, extra parallel writers raise the odds that two agents edit from the same snapshot and one change disappears. The Antigravity team already warns that even with idle waits, parallel swarms chew quota much faster than a normal chat. Five weeks of internal dogfooding went into multi-package migrations and cross-subsystem hunts, which is the right shape of work, and a bad default for a ticket that needs one clean patch.
Across 16 real bug fixes in our 14,000-test codebase, Carbon achieved a 94% resolution rate with zero regressions.
Baki Celtik, AI Engineer, HubX
Kit Langton, a software engineer at OpenCode, said the team handed Carbon a UI-consistency brief, closed the day, and came back to a new Rust UI library and a rewritten desktop app. That is Carbon used as a planner on a hard overnight job, which is the use Google is actually selling.
The Same 13-Day Fairwind Window
Carbon’s public opening came 13 days after its October 28 Fairwind debut. Argon followed the same clock. Google announced Argon for trusted defenders on September 30 and Argon first went to vetted defenders before the October 13 public API drop, also 13 days later. The pattern is now the release method: park the sharpest cyber tools with the 650-plus Fairwind cohort, tune the brakes, then sell a public twin that shares a price card with last month’s model.
On the public twin, Google says high-severity misalignment alerts in live internal coding fell by almost 80%, unsupported claims fell by about 2x, overreach fell 26%, and under-service fell 29%. Successful cheating on Terminal-Bench 4.0 fell to 0.0%, with cheat attempts at 0.4%. Those figures describe a model Google is willing to put on a paid endpoint. They do not describe the Fairwind build that posted the 88.2% find rate with the guardrails off.
Ultra Starts at $99.99 a Month
Consumer access is narrower than the developer API. Carbon is rolling out to Google AI Ultra subscribers in the Gemini app, AI Mode in Search, Antigravity, and AI Studio. Google AI Ultra starts at $99.99 a month for the 5x tier. Google AI Pro, at $19.99 a month, keeps Argon and does not get Carbon. Free users are not in this release, and Google has not set a date for them.
WHO CAN CALL CARBON NOW
- Paid API: Gemini API in Google AI Studio and Android Studio, plus Gemini for business.
- Google AI Ultra: Gemini app, AI Mode in Search, Antigravity app and CLI, and AI Studio.
- Scientists: An application for six months on Antigravity’s highest usage tier, free, subject to eligibility and capacity.
- Cyber defenders: Fairwind only, for the version scored on SEC-bench Pro, ExploitGym, and CWE-bench.
Enterprises can also reach Carbon through Gemini for business and through specialized agents such as CodeMender. The people who can spend $99.99 a month, or who already pay API bills, are the ones Google is inviting to try the heavier model. The people doing daily feature work are the ones Kavukcuoglu just pointed back to Argon.
The $2 and $10 rates last through November 30. On December 1 they double, and every extra thinking token Carbon spends on a job Argon could have finished will cost twice as much.
-
NEWS3 months agoBoston Scientific Cyberattack Leaves Hospitals Waiting on Implants
-
GAMING4 weeks agoXbox Sells Cloud Playtime Packs Ahead of November Caps
-
NEWS3 months agoGuizhou Cave Cafes Spread When Each Cavern Fills
-
GAMING3 months agoThe Second Ned Stark Kit Lands When a Banner Dies
-
NEWS2 months agoSatellites Saw Nepal’s Ice Slope Move Before the Flood
-
NEWS3 months agoOpenAI’s Jalapeño Chip Turns Watts Into Faster Agent Steps
-
NEWS3 months agoMLB Quiet Periods Copy a Freeze Other Leagues Use
-
NEWS3 months agoRoman Space Telescope Launches After a Decade-Long Budget Bet
