Google DeepMind shipped Gemini 4 Carbon on Wednesday, barely two weeks after it released Gemini 4 Argon, and the new model immediately takes the top spot on the Artificial Analysis Intelligence Index v5 with a score of 59, up from Argon's 50.
The pace is unusual even by this industry's standards. Google says Carbon was built for "much harder, long-horizon engineering problems" that need an AI system to keep working through a task for days without losing the thread. Internal teams reportedly ran the model continuously for multiple days, during which it worked through dead ends on its own and tackled root-cause analysis and subtle bug fixes in large, low-level codebases, according to Google.
The gains show up across independent evaluation platforms, not just Google's own claims. On WebDev Arena, Carbon climbed 89 Elo points to 1,769, landing fourth and just two points behind Claude Sonnet 5.5. In Arena.ai's Text Arena, Carbon edged past Argon to take the lead at 1,528 points. On Vals AI, it placed second in general accuracy at 70.36 percent, trailing Claude Sonnet 5.5 by less than half a point, and it topped both the Finance Index and the Legal Research benchmark there, though legal research tasks now average more than 50 minutes each.
Security is the other headline. On Gray Swan's indirect prompt injection test, Carbon's attack success rate dropped to 0.0 percent from Argon's 0.7 percent, meaning every tested manipulation attempt failed, though Google is careful to note that doesn't mean prompt injection as a category is solved. Running without safety guardrails for internal testing, Carbon scored 88.2 percent on SEC-bench Pro, a benchmark for autonomous vulnerability discovery, and led all tested models there. It also topped CWE-Bench for writing fixes.
The improvements aren't free of trade-offs. Hallucination rate rose from 4 percent under Argon to 11 percent under Carbon, still the second-lowest in the Artificial Analysis ranking but a notable jump. The model also burns through more compute: 2.3 times the reasoning tokens and 64 percent more output tokens than Argon on the Intelligence Index, and tasks on the Finance and Tax Agent benchmark now take more than twice as long.
On cost, Carbon lands close to Argon's footprint at $7.76 per task on the Intelligence Index, about 23 percent cheaper than Claude Opus 5.5 at its max setting and roughly 20 percent pricier than GPT-6 Astra. Google is keeping the introductory API pricing unchanged at $2 per million input tokens and $10 per million output tokens, with a 95 percent discount on cached inputs.
Access, though, isn't opening wide right away. Google is rolling Carbon out gradually and has enrolled it in the U.S. Government's pre-release testing process, citing its security capabilities. Selected cybersecurity experts get access this week, with broader access to trusted partners following through DeepMind's Fairwind Program. Developers, businesses and everyday Gemini users will have to wait until Google has gathered more feedback and tuned the safety guardrails before Carbon reaches product interfaces generally.
ADFiled by The AI Desk
Models, assistants and the companies and chips behind them, reported from what was released and what was claimed, with the difference kept clear.
More from this desk →
Be the first to comment
Join the argument. No password, just your email or a passkey.