Google Announces Gemini 4 Carbon, a Frontier Model That Goes to Cyber Defenders First

Google DeepMind on October 28, 2026 announced Gemini 4 Carbon, a frontier model it calls its strongest yet for cybersecurity and long running tasks, and is giving it only to a subset of cyber defenders for now. Artificial Analysis, an independent company that benchmarks AI models, scores Carbon at 59 on its new Intelligence Index v5, one point above Claude Opus 5.5.

Carbon arrives 15 days after Google released Gemini 4 Argon to paying users on October 13, and it follows the same path: Argon was announced on September 30 and first went only to cyber defenders. It is the second Gemini 4 model in a busy month of frontier model releases.

Who can use Gemini 4 Carbon, and when?

Gemini 4 Carbon goes to a subset of cyber defenders starting the week of October 28, 2026, and Google said it will then phase in more trusted partners through its Fairwind Program, according to Google's announcement. Developers, businesses and consumers get it "as soon as possible," after Google gathers feedback and iterates on guardrails.

Google said it is taking part in the US government's voluntary process for testing models before release. It gave no date for a wider rollout, and Ars Technica notes that Google makes no promises on timing. Google describes Carbon as built for "sustained agentic workflows" and says internal teams have seen it run for days on multi step engineering work, such as tracing and patching subtle bugs across large codebases.

What will Gemini 4 Carbon cost?

Gemini 4 Carbon will launch at Argon's introductory price of $2 per million input tokens and $10 per million output tokens, with cached input 95% cheaper than regular input, Google said on October 28, 2026. After the introductory period the price becomes $4 and $20.

For Argon, the discount runs until November 30. For Carbon, Google declined to give an end date when Ars Technica asked.

The standard rate puts Carbon level with Claude Opus 5.5 per token: Artificial Analysis lists Opus 5.5 at the same $4 and $20. Per task the two differ. At standard prices, Artificial Analysis measured Carbon at $7.76 per Intelligence Index task, about 23% less than Opus 5.5 at $10.04 and about 1.2 times GPT-6 Astra at $6.25, because Carbon produced 29% fewer output tokens than Opus 5.5.

How does Gemini 4 Carbon compare with Claude Opus 5.5 and GPT-6 Astra?

In the benchmark table Google published on October 28, 2026, Gemini 4 Carbon has the top score on 15 of 17 rows. GPT-6 Astra leads on FrontierSWE 2 and Claude Opus 5.5 on OSWorld-2.1, and Carbon scores above its predecessor Argon on all 17 rows.

The margins are often small: Carbon is within one point of the best rival on three rows: the AA Index, the Vals Index and SWE-sweep. The table leaves out several current models, including OpenAI's GPT-6.1 Sol, Meta's Muse Wonder Full and Moonshot AI's Kimi K3.1 Pro.

That omission matters on at least one row. Artificial Analysis lists Kimi K3.1 Pro as the leader on Terminal-Bench 4.0 at 67% in its v5 announcement, and its own run put Carbon at 64%, below the 65.2% Google reports. Artificial Analysis credits Carbon with the top score on four of the nine evaluations in the index: HLE-Diamond, Terminal-Bench-Science, AA-Omniscience and AutomationBench-AA.

The independent ranking is also new. Artificial Analysis published Intelligence Index v5 on October 27, 2026, a day before Carbon, with Claude Opus 5.5 on top at 58. On its Coding Agent Index, Carbon in Google's Antigravity CLI scores 59, behind Claude Code with Opus 5.5 at 62. Carbon's hallucination rate on AA-Omniscience rose to 11% from Argon's 4%, still the second lowest of any model Artificial Analysis has measured.

What did Google say about Carbon's safety and security work?

Google said on October 28, 2026 that Carbon uses the same layered safety stack as Argon and was trained further to avoid shortcuts: successful cheating on Terminal-Bench 4.0 fell to 0.0% and cheating attempts to 0.4%. In internal coding work, high severity misalignment alerts dropped by almost 80%, the company said.

Google also reports that Carbon makes about half as many unsupported claims as Argon, with 26% less overreach and 29% less under service on multi step plans. On Gray Swan's indirect prompt injection benchmark, a test of whether hidden instructions in content can hijack a model, Google's chart shows a 0.0% attack success rate for Carbon, against 0.7% for Argon, 1.0% for Claude Opus 5.5 and 8.5% for GPT-6 Astra. The Decoder notes that this result does not mean prompt injection is solved.

On security work itself, Google says Carbon generates secure code more than twice as often as Argon on internal tests, and that its Chrome team found Carbon's patches more than 10 times better than those of "another lab's large frontier model," which it did not name. The Decoder points out that these internal results cannot be independently verified.

What we don't know yet

• When developers, businesses and consumers get access, and through which Google products first.
• When the introductory $2 and $10 price ends.
• Which US government body is testing Carbon, and what that testing covers.
• How Carbon scores against GPT-6.1 Sol, Muse Wonder Full and Kimi K3.1 Pro on Google's own benchmark set.