NEWS
Gemini 4 Carbon Is Live, and Argon Stays the Default
Google opened Gemini 4 Carbon for paying developers and Ultra users, left Argon as the default, and will double API prices on December 1.
Google opened Gemini 4 Carbon to paying developers and AI Ultra subscribers on November 10, and still says most work should stay on Argon. The new model matches Argon’s launch token price, thinks longer on the same job, and the rate doubles after November 30.
Koray Kavukcuoglu, SVP of Google DeepMind and Google’s chief AI architect, spent the launch making Gemini 4 Carbon available while drawing a hard line under it. Carbon is for the jobs that “test the limits of a model’s intelligence.” Argon remains, in his words, “the ideal choice for the majority of software engineering and knowledge work.”
Google Says Most Work Should Stay on Argon
That split is the product, not a footnote. Carbon is the stronger model on independent tests of hard reasoning, science, and long agent runs. Google still routes everyday coding and office work to the sibling it shipped in October, because Carbon spends more thinking tokens per turn and sits in a slower first-token loop.
The company account put it in plainer language the same day: if Argon is great for complex tasks, Carbon is the model you bring in for your hardest problems.
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
Carbon reached paying developers through the Gemini API in Google AI Studio and Android Studio, and through Gemini for business. On the consumer side it is rolling out only to Google AI Ultra, a plan Google prices at about $100 a month, inside the Gemini app, AI Mode in Search, and Google Antigravity. Google AI Pro keeps Argon. Free and Plus accounts do not get Carbon at all.
The first public demos were not chat replies. They were overnight 3D reconstructions, a rebuilt 1780s fortepiano, and a pelican riding a bike across the Mountain View campus, the kind of long agent run that drains quota even when idle workers sit at zero tokens.
The Same Sticker Price, a Higher Bill
Google posted introductory API token prices that copy Argon exactly: $2 per million input tokens and $10 per million output tokens through November 30, with cached input at 95% off. On December 1 both rates double, to $4 and $20. Shops that model the launch sticker as a unit cost will miss the bill that starts that Monday.
Google is explicit that the same price per token is not the same price per task. Carbon “uses more tokens than Argon to explore thoroughly, reason deeply, and run extended verification loops,” Kavukcuoglu wrote. Its default posture is to keep checking. That is the point of the model, and it is also how a $2 list price turns into a larger invoice.
CARBON VS ARGON AT THE METER
| Measure | Carbon (high) | Argon (high) |
|---|---|---|
| API input / output through Nov. 30 | $2 / $10 per 1M | $2 / $10 per 1M |
| API input / output from Dec. 1 | $4 / $20 per 1M | $4 / $20 per 1M |
| Cost per Intelligence Index task | $7.76 | $9.95 |
| Output tokens per index task | 123k | 116k |
| Output tokens to run the full index | 205M | 113M |
| Time to first answer token | 66.85 seconds | 37.58 seconds |
Artificial Analysis prices those index runs at the $4 / $20 rate, so the per-task column is not a launch-week discount. Carbon comes out slightly cheaper per scored task on that board even though it emits more tokens, because the mix of cache hits and input differs. The full-index dump tells the other half: Carbon burns 205M output tokens to run the suite, against 113M for Argon, and the all-in cost rises to $8,240 from $7,620. Per-task thrift and a heavier full-suite burn can both be true, and a swarm of Carbon workers will follow the heavier path.
Latency is the part no spreadsheet hides. Carbon takes 66.85 seconds to its first answer token, against 37.58 seconds for Argon, then finishes a single-request workload in 156.06 seconds against 107.95. Once it is talking it is quicker on the wire, at 80 tokens per second against 68. The wait is in the think, which is exactly what Google is selling.
/swarm Turns One Prompt Into a Crew
Google is previewing /swarm in Antigravity as the way to use Carbon without putting it on every file. One prompt becomes a coordinator, a set of workers in isolated git worktrees so edits do not collide, and adversarial verifiers that test and challenge every diff. Agents post to a shared, append-only message board and park at zero token cost when idle.
WHAT A SWARM ACTUALLY RUNS
- Coordinator: A lead agent reads the brief, splits the work, and keeps the board coherent.
- Workers: Specialists edit in isolated git worktrees so parallel changes never overwrite each other.
- Verifiers: Adversarial testers run the suite and challenge diffs instead of rubber-stamping them.
- Board: An append-only feed every agent can read, with a live roster and a graph you can scrub.
- Idle park: Waiting agents sit at zero tokens until another agent needs them.
The Antigravity team’s own routing advice matches Kavukcuoglu’s. Pin Carbon to coordinator, architecture, cybersecurity audit, and adversarial verification. Leave parallel file edits and routine implementation to lighter, faster models, including Argon. Inside Google, engineers spent five weeks running swarms on multi-package migrations, cross-subsystem root-cause hunts, and architectural redesigns. The same post warns that even with idle waits and hot-swapping, parallel multi-agent systems consume quota much faster than a standard chat.
That is the honest product design. Carbon is not the model you set as default in the picker and forget. It is the expensive chair at the head of a table of cheaper ones.
Why Carbon Guesses More Than Argon
On the live side-by-side of Carbon against Argon on the index, Artificial Analysis scores Carbon (high) at 60 on Intelligence Index v5.0.1 and Argon (high) at 50. Carbon also leads AA-Omniscience at 54 against 46, Terminal-Bench 4.0 at 65.2% against 57.1%, FrontierCode 2.0 at 40.6% against 25.7%, HLE-Diamond at 69.1% against 56.7%, and Terminal-Bench-Science at 64.8% against 52.9%. AutomationBench-AA is a near tie, 77.8% against 77.5%. Both models share a 1M-token context window.
The knowledge test is the row that cuts against the “just smarter” story. Carbon answers 58% of Omniscience questions correctly, up 10 points from Argon’s 48%. Its hallucination rate on that test, the share of non-correct answers that are confident mistakes rather than abstentions, rises from 4% to 11%. Argon still refuses more often. Carbon attempts more, gets more right, and also gets more wrong when it does not know.
WHERE THE INDEX AND THE GUESSES DIVERGE
- Index lead: Carbon (high) 60, Argon (high) 50, a 10-point gap on Artificial Analysis v5.0.1.
- Omniscience: Carbon 54 against Argon 46, with accuracy 58% against 48%.
- Hallucination: 11% for Carbon against 4% for Argon on the same knowledge test.
- Google’s own claim: Carbon ranks first on CursorBench 4.0, Perplexity’s WANDR, and Zapier’s AutomationBench at more than twice lower cost per task than the next best outside model.
Those two pictures can sit together. Omniscience punishes a wrong guess and scores an “I don’t know” at zero, so a model that answers more questions can lead the net score while hallucinating more often than its quieter sibling. Google’s internal write-up says Carbon cuts unsupported claims by about 2x in its own agent tests, which is a different pile of work than AA’s 3,000-question set. On Gray Swan’s indirect prompt-injection benchmark, Carbon’s attack success rate is 0.0%, against 0.7% for Argon. Successful cheating on Terminal-Bench 4.0 is 0.0%, with attempts at 0.4%. High-severity misalignment alerts in live internal coding dropped by almost 80%, Google says, with overreach down 26% and under-service down 29%.
Google also says Carbon ranks second on Artificial Analysis’s Coding Agent Index, behind Claude Opus 5.5. The new Multimodal Knowledge Agent test, built with Artificial Analysis, has Carbon at No. 2 and more than 6x lower average cost per task than the top model. Frontier on a blended index is not frontier on every desk.
Fairwind and the Locked Cyber Build
The version that hunts exploits is not the version in the public API. Google announced Carbon on October 28 and gave it first to a subset of cybersecurity defenders through its Fairwind Program. The public model now in AI Studio is the general-use build. Carbon for cyber, with fewer brakes on offensive work, still moves through the Fairwind Program for cyber defenders.
On the unguarded cyber benches Google published in October, Carbon scored 88.2% on SEC-bench Pro for finding vulnerabilities, 54.8% on ExploitGym for proving they can be exploited, and 71.0% on Collinear’s CWE-bench for patching. The Chrome team said Carbon produced high-quality patches more than 10 times better than another lab’s large frontier model. Internal tests showed more than a 2x lift over Argon at generating secure code. Google’s public build, the company says, blocked 99% of attack attempts in its own red-team tests.
In Project Fortify, Google’s scanning, triage, and repair loop, Carbon hit 100% recall on scanning and 95% precision on triage on an internal benchmark, with more than a 25% relative lift on Argon’s already strong recall against open-source repositories. Google’s Cloud Vulnerability Research team used it to prove two bugs in a popular open-source project were exploitable, on the order of the PHP2Shell case, and to push maintainers to re-class them as critical. Wiz, scanning two large SaaS providers, says the model found 0-day vulnerabilities in under two hours and helped get them fixed.
Chris Corde, chief product officer at SentinelOne, said every model his team tested found the same core memory-corruption bugs in a kernel driver, and that Carbon for cyber went further, separating findings that hold up from those that do not and uncovering four additional verified vulnerabilities. That skill set is why Fairwind exists, and why it is still the gate.
Who Can Use Gemini 4 Carbon Now
Access is a ladder, not a switch. The API opened on November 10 for anyone on the paid Gemini tier. The app did not. Ultra is the consumer tollbooth, and the cyber build is a separate application.
THE ACCESS CALENDAR
- October 28, 2026: Google announces Carbon and starts a Fairwind cohort of cyber defenders while it tunes guardrails.
- November 10, 2026: The paid Gemini API, Gemini for business, and Google AI Ultra surfaces get the general-use model; /swarm enters preview in Antigravity.
- November 30, 2026: The $2 / $10 introductory token rate ends.
- December 1, 2026: Standard API pricing of $4 / $20 applies to Carbon and Argon alike.
The calendar is short on purpose. Anyone who treats the November rate as the lasting cost of running Carbon at high thinking will be wrong by the first week of December.
WHO IS ON THE MODEL, AND WHO IS NOT
- Paid API: gemini-4-carbon in Google AI Studio and Android Studio, plus Gemini for business.
- Ultra: Gemini app, AI Mode in Search, Antigravity (app, CLI, and IDE extensions), and AI Studio, on a plan of about $100 a month.
- Enterprises: Model picker access plus specialized agents such as CodeMender and AlphaEvolve.
- Scientists: An application for six months on Antigravity’s highest usage tier, Argon and Carbon, free, subject to eligibility and capacity.
- Cyber teams: Fairwind, still, for Carbon for cyber.
- Pro, Plus, and free: No Carbon. Pro keeps Argon.
The $100 Ultra door is the argument consumer users are actually having. Carbon is not a free upgrade inside an existing Pro plan. It is a new SKU on a plan Google already sold as a 5x usage bump, with a higher 20x tier above it. Developers who only need the API can skip the subscription and pay tokens. Everyone else is buying a seat.
Two 0-Days in Under Two Hours
The cleanest way to see what Google thinks Carbon is for is to look at the jobs it brags about, not the index. HubX ran it against 16 real bugs in a 14,000-test codebase.
Gemini 4 Carbon treats legacy software contracts with absolute respect. 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 at OpenCode said his team handed Carbon a list of UI requirements at the end of the day and, overnight, got a new Rust UI library and a rewritten desktop app with working demos. Hebbia’s Shanks Salwan called it the strongest model the firm has evaluated for deep financial research, cleaner Excel models, and sharper decks. Cartwheel’s Andrew Carr said six days of light steering improved a tuned cloth simulation in speed and fidelity. Box’s Yashodha Bhavnani said Carbon scored 100% on public-sector tasks in the firm’s tests.
Google’s own Antigravity reels are the same shape of work: a 906K-triangle reconstruction of the Platform 37 office from seven planning PDFs across 10 overnight loops; a Stein-Schule fortepiano from about 1780, 61 keys, rebuilt from CAD and X-rays and dropped into a model of Koerner Hall with seat-specific acoustics; an open-world campus game in which a pelican rides a GBike. Those are coordinator jobs with verifiers, not a coding autocomplete.
WebDev Arena has Carbon at 1,768 Elo, up 89 points, in fourth. That is a real gain and not a crown. The model Google wants on most engineering desks is still Argon. Carbon is in the API, on Ultra, and at the head of a swarm, with the exploit-grade build held for Fairwind and a six-month scientist pass sitting behind a form. The public model is for sale. The version that proved those two SaaS 0-days in an afternoon is not.
-
BUSINESS3 months agoWarsh Rejects Rate Guidance and Still Moves Markets
-
ENTERTAINMENT4 weeks agoDavina McCall Reveals Cardiovascular Disease After Cholesterol Warnings
-
BUSINESS2 months agoDana-Farber Exits an MGB Medicare Advantage Network Early
-
NEWS3 months agoNASA Launches the Roman Space Telescope’s Cosmic Bet
-
NEWS3 months agoOpenAI’s Cyber Letter Puts the Defense Bill on Governments
-
NEWS3 months agoRussia Recycles Its Old Warning Over Storm Shadow Plants
-
ENTERTAINMENT2 months agoPrimetime’s $2.7 Million Preview Tops Pitt’s Costlier Film
-
BUSINESS3 months agoJet Drones Lock Down Kyiv and Strip Kherson of Heat
