Gemini 4 Carbon: what changed, who can use it, what it costs

Google DeepMind has announced Gemini 4 Carbon, two weeks after it launched Gemini 4 Argon. Carbon scores 59 on version 5.0 of the Artificial Analysis Intelligence Index, against Argon's 50, which puts it first overall. Google says it designed Carbon for "much harder, long-horizon engineering problems" that need sustained agentic workflows, and that internal teams ran it for days on single tasks such as root-causing and patching subtle bugs across large codebases and low-level systems.

The Verge reports that Carbon outperforms the top models from OpenAI and Anthropic on some benchmarks. The detailed results show where that holds and where it does not.

Where Carbon beats Argon and where rivals still lead

According to Artificial Analysis, Carbon at its highest reasoning tier leads the index mainly on complex reasoning, knowledge retrieval and scientific tasks, taking first place in four of the nine individual evaluations. In agentic knowledge work on AA-Briefcase, Carbon reaches 1,609 Elo, 126 points above Argon, driven by presentation quality and analytical quality. It still trails Kimi K3.1 Pro, Muse Wonder and Grok 4.7 on that test. In web development, Carbon gained 89 Elo points to reach fourth place on WebDev Arena. On Vals AI, Carbon scores 70.36 percent accuracy, second behind Claude Sonnet 5.5 at 70.85 percent. It leads the Vals AI Finance Index and Legal Research benchmark, but Finance and Tax Agent tasks take more than twice as long as with Argon, and Legal Research tasks average more than 50 minutes each. Carbon generates 2.3 times as many reasoning tokens and 64 percent more output tokens than Argon, though still less than half the output of Sonnet 5.5 and Opus 5.5. Answer accuracy rises from 48 to 58 percent compared to Argon, but the hallucination rate climbs from 4 to 11 percent, which Artificial Analysis still ranks as the second lowest. In agentic coding on Antigravity, Carbon jumps from 53 to 59 but still trails Claude Code with Opus 5.5 by three points.

Carbon's cybersecurity capabilities and the Fairwind rollout

Cybersecurity is where Google is most explicit about Carbon's advances. On SEC-bench Pro, which measures vulnerability discovery, Carbon scores 88.2 percent. On ExploitGym, which measures turning vulnerabilities into working exploits, it scores 54.8 percent. On CWE-bench, which covers patching, it scores 71.0 percent. Google ran these benchmarks without guardrails and computed the SEC-bench Pro and ExploitGym scores itself. On Gray Swan's indirect prompt injection benchmark, the attack success rate fell from 0.7 percent with Argon to 0.0 percent. The Decoder notes that this does not mean prompt injection is solved, and that the test appears to be saturated. Google also reports that Carbon produces secure code more than twice as often as Argon on internal benchmarks, and that its Chrome team found Carbon more than ten times better than another lab's frontier model at writing high-quality patches. These internal results cannot be independently verified. Wiz reports Carbon leading its web penetration testing benchmarks. Access reflects the security focus: Carbon is available first to a subset of cyber defenders. Access then expands to more trusted partners through the Fairwind Program, which prioritises governments, critical infrastructure operators and core technology platforms. Developers, businesses and consumers get Carbon later, with no date given.

Gemini 4 Carbon pricing on the API

Carbon launches at the same introductory price as Argon: $2 per million input tokens and $10 per million output tokens, with cached input tokens priced at 95 percent off the input price. According to The Decoder, this is a 50 percent introductory discount that runs through the end of 2026. After it ends, prices rise to $4 per million input tokens and $20 per million output tokens. At those list prices, Artificial Analysis puts Carbon's cost at $7.76 per task on its index, about 23 percent less than Claude Opus 5.5 at $10.04, roughly 1.2 times GPT-6 Astra at $6.25, and almost exactly Argon's $7.68.

What this means for businesses comparing frontier models

For a business weighing Gemini 4 Carbon against Claude Opus 5.5 or GPT-6 Astra, the picture is mixed. Carbon leads the Artificial Analysis Intelligence Index overall, leads legal research and finance on Vals AI, and leads the cybersecurity benchmarks Google reported. But Claude Sonnet 5.5 still edges it on overall Vals AI accuracy, Claude Code with Opus 5.5 still leads agentic coding on Antigravity, and complex tasks take considerably longer than with Argon. The hallucination rate rising from 4 to 11 percent matters for any workflow where factual precision counts. Many of the headline security figures are Google's own or self-run, with no independent replication yet. Most practically, unless a business is a vetted cyber defender or Fairwind partner, Carbon is not an option right now. For those who can get access, costs will double per token once the introductory period ends.