Gemini 4 Carbon is out: what changed, what it costs, and how it compares to Claude

What Google released

Gemini 4 Carbon is the second model of the Gemini 4 family, after Gemini 4 Argon. Google first gave it to a group of cybersecurity defenders through its Fairwind Program two weeks ago; today it opens to developers, enterprises, scientists and consumers (launch post). The model card says Carbon is based on Argon and its knowledge cutoff is March 2026.

• →Developers: the Gemini API paid tier in Google AI Studio and Android Studio, model id gemini-4-carbon.
• →Enterprises: Gemini for business, through the API or the model menu, plus Google Antigravity and CodeMender.
• →Everyday users: the Gemini app and AI Mode in Search, for Google AI Ultra subscribers.
• →Scientists: an application program with six months of free access to Argon and Carbon in Google Antigravity.
• →Security teams: a cyber version through the Fairwind Program, for verified defenders.
• →A preview of /swarm in Google Antigravity, which turns one prompt into a team of agents you can follow live.

Gemini 4 Carbon price

Two details change the real bill. Thinking tokens are billed as output, and Carbon thinks more than Argon on every turn, so the same prompt costs more on Carbon at the same per-token price. And the introductory price ends on November 30: whatever you test this month will cost twice as much per token in December.

Gemini 4 Carbon benchmarks

Google's model card compares Carbon with Argon, OpenAI's GPT-6 Astra, Claude Opus 5.5 and Claude Fable 5.1 (not 5.2, which came out the day before this launch). Results are as of October 2026. On a phone, the table scrolls sideways.

Carbon tops 15 of these 17 rows. The two it does not win are one coding test and one computer-use test: GPT-6 Astra leads FrontierSWE 2 (65.5% vs 61.2%) and Claude Opus 5.5 leads OSWorld-2.1 (87.2% vs 81.5%). The widest gap is long context: on GraphWalks, run at 256k to 1M tokens of input, Carbon scores 96.2%, at least 12 points ahead of every other model in the table.

Gemini 4 Carbon vs Claude Fable 5.2

Google's table stops at Fable 5.1. Anthropic's Fable 5.2 launch table, published a day earlier, does include Carbon. Where both models have a score:

Fable 5.2 wins six of the seven. Carbon wins AutomationBench, where an agent has to finish business workflows across apps, and comes within one point on reading charts.

The independent check

Artificial Analysis runs its own tests. Its Intelligence Index had Carbon at number one when the model was announced on October 28. Since Fable 5.2 arrived, Carbon (high) sits fourth of 74 with a score of 60, behind Fable 5.2 at max (65), xhigh (62) and high (62) effort.

Two things stand out. Carbon costs less per task than Fable 5.2 at high effort ($7.76 vs $9.08) for two points less, and while the introductory price lasts its tokens cost half of what that figure assumes. But it is slow and wordy: 25.6 minutes per test task against a 12.8 minute median, and 205 million output tokens to run the index against a 72.8 million median. On the AA-Omniscience knowledge test it scored highest at its announcement, with an 11% hallucination rate, the second lowest of any model at the time (Artificial Analysis).

What Hacker News says

The Hacker News launch thread had 233 points and 125 comments when we read it, three hours after release. The top complaint was access: readers on Google plans below Ultra still could not pick Carbon in the Gemini app. Others noted it uses more tokens than Argon and arrived a day after Fable 5.2, which several still preferred.

What Gemini 4 Carbon is best at

Google positions Carbon for long jobs that need deep checking. In its words, Argon remains the choice for most software and knowledge work, while Carbon is built for complex multi-step coding and debugging and open-ended research, using more tokens to explore, reason and run extended verification loops. The examples in the launch post:

• →Real work for less: Google says Carbon ranks first 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.
• →Dense documents: on the new Multimodal Knowledge Agent benchmark, built with Artificial Analysis, it ranks second at more than 6x lower average cost per task than the top model.
• →Coding: HubX reports a 94% resolution rate with zero regressions across 16 real bug fixes in a 14,000-test codebase, and OpenCode says Carbon built a new Rust UI library overnight.
• →Security: Google says it leads on finding, exploiting and patching vulnerabilities, and in Google's Project Fortify it reached 100% recall on scanning and 95% precision on triage on an internal benchmark.
• →Science: Google researchers used it to help prove security properties of the code behind Signal's cryptographic protocol, and to find 40 times more related protein sequences on average across 7 hard protein families from the CASP17 test set.
• →Honesty: on Google's 375-task agentic honesty benchmark it passed 92% of tasks, vs 70% for Argon.

Gemini 4 Carbon vs Argon: when to use which

Both share the 1M token window, the tools and the per-token price. The difference, per Google's developer guide, is how many thinking tokens each spends per turn. In practice:

• →Use Carbon for root-cause debugging across a big codebase, architecture work, reports built from many dense PDFs or spreadsheets, proofs and security audits.
• →Keep Argon for everyday feature work, interactive coding and fast loops: fewer thinking tokens per turn and faster answers.
• →Start Carbon at the default medium thinking level. Google recommends testing low, medium and high and keeping the lowest level that meets your bar.
• →If you already pay for Claude, Fable 5.2 still scores higher on most shared tests. Carbon is the one to try for long-context and business-workflow jobs, and this month it costs half its list price.

How to try Gemini 4 Carbon

• →Google AI Studio: pick Gemini 4 Carbon in the model list (paid tier).
• →Gemini app and AI Mode in Search: with a Google AI Ultra subscription.
• →API: model id gemini-4-carbon on the Gemini API paid tier, or through Gemini for business.

A minimal call with the Google Gen AI Python SDK, following Google's quickstart. The thinking level can be low, medium (the default) or high.

from google import genai

client = genai.Client()

interaction = client.interactions.create(
    model="gemini-4-carbon",
    input="Find the root cause of this flaky test and propose a fix.",
    generation_config={"thinking_level": "medium"},
)

print(interaction.output_text)

What it means for you

A stronger model does not fix the usual reason AI builds go wrong. Carbon can hold a million tokens in context, and your AI can keep notes and memory files, but nothing makes it read them before it acts, so it repeats what you already ruled out and redoes work that was finished. A bigger window holds more; it does not decide what gets checked first.

That is why we built Journals DNA: your AI has to stop and check its memory before it acts, then it builds on what you already decided, and any AI or teammate can pick up the same file and continue. This post was made with Motion DNA, and everything else we build is at vibedna.ai.