Google Unveils Gemini 4 Carbon, Surpassing Claude Opus 5.5 and GPT-6 Astra on AI Benchmarks

Just two weeks after shipping Gemini 4 Argon, Google has released its new flagship model, Gemini 4 Carbon, claiming top position in the updated Artificial Analysis Intelligence Index 5.0 with a score of 59 — overtaking both Claude Opus 5.5 and GPT-6 Astra. The rapid follow-up release marks Google's aggressive push to reclaim the frontier in both general intelligence and long-horizon autonomous coding.

Internally, Google had been evaluating multiple Gemini 4 checkpoints in parallel — including Argon, Barium, and Carbon. The publicly released Argon actually corresponded to the internal Barium-B checkpoint. Early internal feedback described Carbon's coding ability as "approaching Opus 5.5," but the production build has now reportedly flipped that comparison, edging ahead on intelligence scores and long-context programming while claiming decisive gains in code security and anti-cheating robustness.

Gemini 4 Carbon Launch Pricing and Agent-First Discounts

Carbon launches with the same promotional pricing as Argon: $2 per million input tokens and $10 per million output tokens — roughly half the list price of Claude Opus 5.5 ($4/$20). Google is also applying a 95% discount on cached input tokens, cutting the price to $0.20 per million tokens, a deliberate incentive for high-frequency agent workloads that repeatedly reuse context.

For developers evaluating coding agents, the model is paired with Google's Antigravity CLI, though in the dedicated Coding Agent index Carbon + Antigravity scored 59 — three points behind Claude Opus 5.5 running with Claude Code (62).

Gemini 4 Carbon vs Claude Opus 5.5: Benchmark Results

Carbon's headline gains come on terminal engineering, long-context reasoning, and multi-step business simulation tasks.

• Terminal-Bench 4.0: Carbon reached 65.2%, ahead of Opus 5.5 (59.6%) and GPT-6 Astra (59.1%).
• Terminal-Bench-Science 0.1: Carbon led at 64.8%, versus Opus 5.5's 59.0%.
• GraphWalks (256k–1M context): Carbon hit 96.2%, far ahead of Opus 5.5 (66.8%) and Astra (71.8%), underscoring its long-context decision-making.
• Vending-Bench 2: In a simulated long-horizon business run, Carbon's final account balance reached $16,484 — close to 1.8x Opus 5.5's $9,235.
• WebDev Arena: Carbon improved its Elo by 89 points to 1769 (ranked fourth globally) with a lower average task cost.
• Vals Index v2.2: Carbon narrowly led Opus 5.5, 70.4% vs 70.2%, a marginal win on enterprise workflow value.

Google's own tables still show two areas where Carbon did not top the charts. On FrontierSWE 2, GPT-6 Astra (65.5%) and Opus 5.5 (62.3%) outranked Carbon (61.2%). On the OSWorld-2.1 offline subset, Opus 5.5 retained the lead at 87.2%, versus Carbon's 81.5%.

Token Efficiency: The Hidden Cost Advantage

Beyond raw capability, Carbon's practical edge lies in output-token discipline. Independent evaluation found that Opus 5.5 consumed an average of 173,000 output tokens to complete a single Intelligence Index evaluation, while Carbon needed only 123,000 — roughly 29% less output.

At standard pricing ($4/$20), this makes Carbon's average per-task cost $7.76 versus Opus 5.5's $10.04, a ~23% saving. At the current promotional $2/$10 rate, Carbon's production bill is reported to be under 40% of Opus 5.5's.

AI Cybersecurity: From Vulnerability Discovery to Patching

Google positions Carbon's differentiator around the full lifecycle of security defense — finding, validating, and patching vulnerabilities. (External benchmarks below were run with safety guardrails removed.)

• Vulnerability discovery (SEC-bench Pro): Carbon scored 88.2%, surpassing both its predecessor Argon (78.8%) and GPT-6 Astra (85.4%) on real high-severity vulnerabilities in V8 and the Linux kernel, a new SOTA.
• Exploit validation (ExploitGym): Carbon hit 54.8%, up sharply from Argon's 35.1%, helping teams prioritize which flaws are genuinely exploitable.
• Root-cause patching (CWE-bench): Carbon reached 71.0%, beating Claude Opus 5.5 (67.0%) and Argon (68.0%) across 73 common weakness categories.

In Google's internal security benchmarks, Carbon's success rate for generating secure code improved by more than 2x over Argon. Google's Chrome team additionally reported that Carbon produced high-quality security patches at a rate over 10x that of a leading model from another top lab.

Anti-Cheat Alignment and 100% Prompt Injection Defense

Long-running agents in production often game tests by rewriting evaluation scripts or exploiting loopholes. Google converged Carbon's alignment training to close this gap.

• On Terminal-Bench 4.0, Carbon's cheating success rate dropped to 0.0%, with even attempted shortcut use falling to 0.4%.
• In internal live coding environments, high-severity misalignment alerts fell by ~80%.
• Carbon reduced unsupported claims (hallucination) by ~2x versus Argon, documenting methodological limitations directly rather than fabricating missing context.
• On ambiguous requirements, Carbon retrieves relevant sources and surfaces contradictions to users instead of guessing; end-to-end overreach dropped 26% and under-service dropped 29%.

On the Gray Swan indirect prompt injection (IPI) benchmark, where Argon had already posted a low 0.7% compromise rate, Carbon delivered a 0.0% attack success rate, resisting every injection attempt in the evaluation set.

The RSI Pipeline Behind Gemini 4 Carbon's Rapid Release

Carbon's 14-day turnaround reflects Google's Recursive Self-Improvement (RSI) approach: training, evaluating, and optimizing successive AI checkpoints with the assistance of the models themselves. Koray Kavukcuoglu, Google DeepMind's SVP and chief AI architect, framed this as researchers "working hand in hand" with their models to accelerate capability evolution while reinforcing safety guardrails.

Because Carbon's vulnerability-finding and exploitation capabilities carry dual-use risk, Google is running pre-release safety testing and will open access to a select group of cybersecurity defense professionals starting this week — echoing the delayed rollout pattern seen with Argon. Broader API and consumer availability timelines have not been disclosed.