Models
Gemini 4 Carbon: Google retakes the lead but keeps it locked away
Gemini 4 Carbon tops Artificial Analysis and LMArena, but it is not yet available to developers or the general public. Scores, pricing, access.
On October 28, Google announced Gemini 4 Carbon, which it calls its strongest model, with state-of-the-art cybersecurity and long-horizon capabilities (Google). It arrives two weeks after Gemini 4 Argon, which it builds on. It also takes first place on two rankings developers follow. But Carbon is not yet available to developers or the general public: only a few cyber defenders have access, and Google gives no release date (Google).
What follows separates the numbers Google published, those from independent evaluators, and our own reading, flagged as such. Prices are in dollars per million tokens.
What Carbon is
Carbon is not a lighter "Flash" model but a stronger version of Argon, built for long engineering tasks. Google writes that its teams have seen it run "for days" to find the root cause of subtle bugs and fix them in large codebases (Google).
The other focus is defensive cybersecurity. Google reports 88.2% on SEC-bench Pro for finding vulnerabilities, 54.8% on ExploitGym for checking that they can be exploited, and 71.0% on CWE-bench for fixing them. In its internal tests, Carbon is more than twice as successful as Argon at writing secure code. The Chrome team reports patches "more than 10x" better than those of another lab's large frontier model, which it does not name (Google).
Google publishes neither the parameter count nor the maximum output length. Artificial Analysis lists a one-million-token context window, with text, image, video and speech input and text output (Artificial Analysis).
First on the Artificial Analysis index
On version 5 of the Artificial Analysis Intelligence Index, overhauled the day before, Carbon takes the lead with 59 points, one more than Claude Opus 5.5.
Artificial Analysis Intelligence Index v5, out of 100
█ Gemini 4 Carbon
░ Gemini 4 Argon
▒ competitors
Gemini 4 Carbon 59 │██████████████████████████████
Claude Opus 5.5 58 │▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒
GPT-6 Astra 57 │▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒
GPT-6.1 Sol 55 │▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒
Gemini 4 Argon 50 │░░░░░░░░░░░░░░░░░░░░░░░░░
Claude Fable 5.1 49 │▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒Sources: Artificial Analysis, Carbon article and Index v5 announcement, October 27 and 28, 2026; Claude Fable 5.1 score from the Google DeepMind table. Carbon at high reasoning, the others at maximum effort.
A one-point gap is thin. Artificial Analysis notes that Carbon leads 4 of the index's 9 evaluations (HLE-Diamond, Terminal-Bench-Science, AA-Omniscience and AutomationBench-AA). It is only second on the Terminal-Bench 4.0 and FrontierCode coding tests. On AA-Briefcase, which measures office work, Carbon gains 126 Elo points over Argon (1,609) but stays behind Grok 4.7 (1,638), Muse Wonder (1,659) and Kimi K3.1 Pro (1,678) (Artificial Analysis).
Reliability is a mixed picture. Carbon has the best AA-Omniscience score (54, against 48 for Opus 5.5), but its hallucination rate rises from 4% for Argon to 11%. That is still far below Claude Opus 5.5 and GPT-6 Astra, both at 50% (Artificial Analysis).
What Carbon brings over Argon and the competition
Google publishes a 17-test table where Carbon beats Argon on every line. Against competitors, the picture is more mixed. Here are five representative rows.
Scores published by Google, in %
█ Gemini 4 Carbon ░ Gemini 4 Argon
▒ GPT-6 Astra ▓ Claude Opus 5.5
Terminal-Bench 4.0 (AA)
Carbon 65.2 │████████████████
Argon 57.1 │░░░░░░░░░░░░░░
GPT-6 Astra 59.1 │▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒
Opus 5.5 59.6 │▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓
FrontierSWE 2
Carbon 61.2 │███████████████
Argon 55.0 │░░░░░░░░░░░░░░
GPT-6 Astra 65.5 │▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒
Opus 5.5 62.3 │▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓
OSWorld-2.1
Carbon 81.5 │████████████████████
Argon 74.7 │░░░░░░░░░░░░░░░░░░░
GPT-6 Astra 80.9 │▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒
Opus 5.5 87.2 │▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓
GraphWalks 256k-1M
Carbon 96.2 │████████████████████████
Argon 84.2 │░░░░░░░░░░░░░░░░░░░░░
GPT-6 Astra 71.8 │▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒
Opus 5.5 66.8 │▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓
ExploitGym
Carbon 54.8 │██████████████
Argon 42.0 │░░░░░░░░░░░
GPT-6 Astra 42.4 │▒▒▒▒▒▒▒▒▒▒▒
Opus 5.5 34.5 │▓▓▓▓▓▓▓▓▓Source: Google DeepMind table; Google's methodology. These are vendor numbers, not an independent evaluation.
The clearest gains are in long context (GraphWalks, between 256,000 and 1 million tokens, where Carbon is well ahead of GPT-6 Astra and Opus 5.5), cybersecurity and science. On the other hand, GPT-6 Astra keeps the edge on FrontierSWE 2, a software engineering test, and Claude Opus 5.5 on OSWorld-2.1, which measures computer use. On SWE-sweep, Carbon (9.6%) beats GPT-6 Astra (9.5%) by only a tenth of a point (Google DeepMind).
For coding agents, Artificial Analysis also tested it inside Google's own tool, Antigravity CLI. The pair scores 59 on its Coding Agent Index, 3 points behind Claude Code with Opus 5.5 (62) (Artificial Analysis).
Google also highlights behavioral safety. On Terminal-Bench 4.0, Carbon no longer succeeds at any cheating attempt (0.0%) and only tries to cheat in 0.4% of cases. High-severity misalignment alerts drop by almost 80% in internal testing. On Gray Swan's indirect prompt injection test, no attack succeeds, against 0.7% for Argon (Google).
First on LMArena, but preliminary
On LMArena, where users compare two answers blind, Carbon also comes first.
LMArena, text leaderboard, Elo score
Bars start at 1,460 points, 1 block = 3 points
█ Gemini 4 Carbon ░ Gemini 4 Argon ▒ competitors
Gemini 4 Carbon 1,528 │███████████████████████
Gemini 4 Argon 1,520 │░░░░░░░░░░░░░░░░░░░░
Claude Opus 5.5 1,509 │▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒
Muse Wonder 1,507 │▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒
Claude Fable 5.1 1,504 │▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒
GPT-6.1 Sol 1,481 │▒▒▒▒▒▒▒
GPT-6 Astra 1,476 │▒▒▒▒▒Source: LMArena, text leaderboard of October 27, 2026, checked October 28. The bars do not start at zero, which magnifies the gaps.
The score is marked "preliminary" and rests on 4,403 votes, against 11,889 for Argon. Its margin of error (±9) covers the 8-point gap with Argon: LMArena places Carbon between 1st and 2nd (LMArena). On WebDev Arena, which covers web development, Google reports an 89-point gain, in 4th place at 1,768 (Google).
Pricing and availability
Carbon is not yet available to developers or the general public: for now Google reserves it for cyber defense partners, with no release date (Google). The prices below are announced prices, not rates you can use today.
Carbon will launch at the same introductory price as Argon: $2 input and $10 output, with cached input tokens 95% off. After that, the price will be $4 and $20 (Google). Google does not say when this introductory period ends, while Argon's runs until November 30 (Google).
| Model | Input / output price | Access |
|---|---|---|
| Gemini 4 Carbon | Announced: $2 / $10 at launch, then $4 / $20 | Not yet available: Fairwind Program only, no public API |
| Gemini 4 Argon | $2 / $10 until Nov 30, then $4 / $20 | Gemini API, Antigravity |
| Claude Opus 5.5 | $4 / $20 | Claude API |
| GPT-6 Astra | $10 / $50 | OpenAI API |
| GPT-6.1 Sol | $2 / $10 | OpenAI API |
Sources: Google, Carbon, Google, Argon, Antigravity, Anthropic, OpenAI. Standard rates for requests under 272,000 tokens at OpenAI. Checked October 28, 2026.
Per-token price is not the whole story. Artificial Analysis measures the cost of a full task in its index, at Carbon's standard $4/$20 rate.
Average cost per Artificial Analysis index task, in $
Standard pricing, without launch discount
█ Gemini 4 Carbon ░ Gemini 4 Argon ▒ competitors
Claude Fable 5.1 13.04 │▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒
Claude Opus 5.5 10.04 │▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒
Gemini 4 Carbon 7.76 │████████████████
Gemini 4 Argon 7.68 │░░░░░░░░░░░░░░░
GPT-6 Astra 6.25 │▒▒▒▒▒▒▒▒▒▒▒▒▒
GPT-6.1 Sol 1.27 │▒▒▒
GPT-6.1 Terra 0.96 │▒▒Sources: Artificial Analysis, Carbon article (Carbon, Argon, Opus 5.5, Astra) and Index v5 announcement (Fable 5.1, Sol, Terra).
Carbon costs about 23% less per task than Opus 5.5 while writing 29% fewer output tokens (123,000 against 173,000). It costs as much as Argon for 9 more points, but about 1.2 times as much as GPT-6 Astra (Artificial Analysis). GPT-6.1 Sol is still six times cheaper per task, for 4 fewer points.
On access, Carbon goes this week to "a subset" of cyber defenders, then to more trusted partners in the Fairwind Program. Google still wants to "gather feedback" and refine its guardrails before opening it to developers, businesses and consumers "as soon as possible", with no date (Google). Fairwind has more than 650 partners, with priority for governments, critical infrastructure operators and core technology platforms. Access is controlled: organizations are vetted, phishing-resistant authentication is required, and access may not be resold or shared (Google DeepMind).
As of October 28, Carbon appears neither in the Gemini API model list, nor in its changelog, nor in its pricing page (Google). Google has announced no date for the Gemini app, Vertex AI, Antigravity or the Gemini CLI.
What it changes for developers
For now, nothing concrete: no API model ID is published and the model is in no public product. The only hint about timing comes from Argon. Announced the same way, it then reached the Gemini API and Antigravity, both the app and the command line, for Google AI Pro and Ultra subscribers (Antigravity). Our guide to Gemini CLI and Antigravity CLI covers these tools.
Our reading: when Carbon arrives, it will matter most for three uses. Very long agent tasks, which Google puts forward. Large contexts, where its GraphWalks lead is clearest. And code security audits, its stated strength. For everyday coding, the gap with Opus 5.5 and GPT-6 Astra is thin on published tests, and GPT-6.1 Sol costs far less per task. The introductory price with no end date is also a risk for anyone building a budget: it will double.
Two things to watch: the date the API opens, and Carbon's LMArena score once it has enough votes to stop being preliminary.
Sources
- Google, Gemini 4 Carbon announcement, October 28, 2026, checked 2026-10-28
- Google DeepMind, Gemini 4 performance table and methodology, checked 2026-10-28
- Google DeepMind, Fairwind Program
- Artificial Analysis, Gemini 4 Carbon is the new number 1, October 28, 2026
- Artificial Analysis, Intelligence Index v5, October 27, 2026
- LMArena, text leaderboard, October 27, 2026, checked 2026-10-28
- Google, Gemini 4 Argon announcement
- Google Antigravity, Gemini 4 Argon in Antigravity
- Google, Gemini API models, checked 2026-10-28
- Anthropic, pricing and OpenAI, API pricing, checked 2026-10-28
Spotted an error in this article? Report it · Our corrections policy
