Google Says Gemini 4 Carbon Tops Cyber Benchmarks as Early Access Stays Restricted
Google’s Gemini 4 Carbon claims state-of-the-art cybersecurity and long-horizon coding, but early access is limited to cyber defenders and the benchmark results are vendor-reported.
Google DeepMind announced Gemini 4 Carbon, a frontier model it says leads on cybersecurity and long-horizon agentic coding. Carbon ranks first on Artificial Analysis Intelligence Index v5.0 at 59 and posts 88.2% on SEC-bench Pro, 54.8% on ExploitGym and 71.0% on CWE-bench, though those evaluations were run without guardrails.
The model builds on Gemini 4 Argon, launched two weeks earlier, and will start with a subset of cyber defenders through the Fairwind Program while Google participates in U.S. pre-release testing. Introductory pricing matches Argon at $2 per million input tokens and $10 per million output tokens, rising to $4 and $20 after the introductory period.
Google reports internal gains in secure code generation, reduced cheating and misalignment alerts, and 0.0% prompt-injection success on Gray Swan’s benchmark.
Key takeaways
- Carbon is built for sustained agentic workflows; Google says internal teams have seen it run for days through dead ends on multi-step engineering tasks such as root-causing and patching bugs across large codebases.
- On WebDev Arena, Carbon gains 89 elo points to reach fourth place at 1768. Google says the Chrome team reported Carbon generated high-quality patches more than 10x better than another lab’s large frontier model.
- Google reports Carbon cut successful cheating on Terminal-Bench 4.0 to 0.0%, with attempts at 0.4%, while high-severity misalignment alerts in internal coding environments fell by almost 80%.
- Carbon reduced overreach by 26% and under-service by 29% in internal testing, and Google says it cuts unsupported claims by about 2x over Argon, putting caveats into deliverables or withholding blocked actions.
- The model will reach developers, businesses and consumers later; Google is gathering feedback on guardrails and says Carbon uses the same multi-layered safety stack as Argon, including prompt-injection resilience and chain-of-thought monitoring.
What people think
Analysis details
40 comments analyzed. Analyzed .
The sample selects active discussion branches and includes available parent comments. It may omit parts of the full thread. The analysis included 40 of 108 usable stored comments because of the input limit. Selected themes do not measure community opinion or how common a view is.
Element names obscure model hierarchy
Argon and Carbon are read as arbitrary elements rather than a clear generation or tier. Carbon replaces Argon at the same price, and the public Argon was reportedly named Barium internally, deepening confusion. Numeric versions such as 4.1 or 4.5 are proposed instead. Comparisons to OpenAI’s Luna/Terra/Sol/Astra and Apple’s versioned releases show naming can encode size or generation, but Carbon and Argon do not.
Access and product experience frustrate
Gemini 4 Argon is limited to Google AI Pro subscribers, leaving Plus subscribers without it, and the Gemini app surfaces older models such as 3.6 Flash and 3.1 Pro. The model selector lists 4 Argon as new, but finding or using it is described as difficult. API setup, batch-mode permissions and low usage limits add friction. Argon’s public availability remains unclear.
Benchmarks arrive before public release
Showing benchmark scores while the model remains restricted is called less meaningful because a stronger competitor may appear before launch. A counterpoint notes that Argon shipped shortly after its announcement, that government evaluation is already underway, and that the timing may be intended to get ahead of leaks and coincide with earnings.
Google’s distribution versus rival mindshare
One argument holds that Google’s products, brand, hardware, data and billion-user distribution give it an advantage over Anthropic, whose customers are mostly enthusiasts. Another counters that Google often abandons products and that its non-core offerings lag competitors, while ChatGPT’s 1.2 billion weekly users show mindshare is not guaranteed. Meta is also mentioned as a possible contender.
Google’s execution and incentives questioned
A cynical view holds that Carbon exists mainly for internal performance reviews rather than users, and that Google’s bureaucracy and product-abandonment record undermine its position. The claim that the model is not aimed at users is challenged as needing proof. The skepticism extends to whether Google can convert research strength into durable products.