Google DeepMind unveils Gemini 4 Carbon, but holds it back for cyber defenders first

Google DeepMind announced Gemini 4 Carbon on Wednesday. The company calls it its strongest model, aimed at cybersecurity work and long-running agentic tasks. It arrives two weeks after Gemini 4 Argon. What sets this launch apart is that almost nobody can use the model yet. Starting this week, access goes only to a subset of cyber defenders, and Google says it is taking part in the U.S. government's voluntary process for pre-release model testing.

What Google announced

The announcement post, signed by Koray Kavukcuoglu, Google DeepMind's SVP and Google's Chief AI Architect, says Carbon is built for sustained agentic workflows. According to Google, internal teams have run it for days on multi-step engineering work such as root-causing and patching subtle bugs across large codebases and low-level systems. Google also says the model improved in full-stack web development, multimodal understanding and scientific reasoning.

Most of the headline numbers come from Google's own post:

• Artificial Analysis Intelligence Index v5.0: first place with a score of 59, a third-party index that Google cites.
• Cybersecurity benchmarks (run without guardrails, per Google): 88.2% on SEC-bench Pro (finding vulnerabilities), 54.8% on ExploitGym (exploitation) and 71.0% on CWE-bench (patching).
• WebDev Arena: up 89 Elo points to fourth place, at 1768.
• Internal results: more than twice Argon's success rate at generating secure code, and a Chrome team report that Carbon wrote high-quality patches more than 10 times better than another lab's large frontier model, which Google does not name.

The internal results cannot be checked from outside. The external benchmarks were selected by Google.

Access and price

Google says it will widen access gradually to additional trusted partners through its Fairwind Program, which it describes as giving vetted partners a head start against cyber threats while models are deployed safely. Developers, businesses and consumers will get Carbon "as soon as possible", in Google's words, after it collects feedback and adjusts guardrails. No date has been given.

Pricing is already set. Carbon launches 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. After the introductory period ends, Google lists $4 and $20. The post does not say when that period ends.

How it looks from outside Google

The Decoder reviewed third-party leaderboards published alongside the launch, and the picture there is more mixed. In its account of Artificial Analysis data, Carbon still trails Claude Opus 5.5 on agentic coding. Its answer accuracy rose from 48% to 58% compared with Argon, but its hallucination rate also rose, from 4% to 11%. The Decoder also reports that on Vals AI's index Carbon scores 70.36%, between Claude Sonnet 5.5 (70.85%) and Claude Opus 5.5 (70.18%), and that it produces about 2.3 times as many reasoning tokens as Argon. We could not open the Artificial Analysis or Vals AI pages from our environment, so these figures come from The Decoder's reporting, not from our own check.

Safety claims

Google puts most of the post's weight on alignment. It says Carbon cut successful cheating on Terminal-Bench 4.0 to 0.0%, with cheating attempts at 0.4%. It also says high-severity misalignment alerts in internal coding work fell by almost 80%, unsupported claims fell by about half compared with Argon, overreach fell by 26% and under-service by 29%. On Gray Swan's indirect prompt-injection benchmark, Google reports a 0.0% attack success rate, down from Argon's 0.7%. These are vendor measurements. The Decoder notes that a zero score suggests the benchmark is saturated, not that prompt injection is solved.

Same-day context from the earnings call

Hours later, on Alphabet's third-quarter call, Sundar Pichai repeated that Carbon is being tested by the U.S. government. He said Google's first-party model APIs now process nearly 32 billion tokens per minute, and that Argon usage on Google's developer platforms ramped three and a half times faster in its first week than Gemini 3 did, even with capacity constraints. He also said more models are coming soon, including more efficient versions for developers who need speed and lower cost.

What it changes for developers today

For now, very little. You cannot select Carbon in the API, so it does not change model choice this week. What you can do is plan for it. The listed price is the same as Argon's, so cost estimates done for Argon carry over, at least until the introductory price ends. The token reports suggest a real catch, though: if Carbon spends far more reasoning tokens per task, the cost of a finished job can rise even when the price per token stays the same. Once Carbon opens up, measure cost per completed task on your own workloads, not just the price sheet.

The rollout also matters beyond Google. A frontier lab is holding back general availability because of the model's cyber capabilities, giving defenders early access and putting it through voluntary government testing first. Whether Google keeps to that order, and how long the restricted period lasts, is the next thing to watch.