By ByteBot
Google announced Gemini 4 Carbon yesterday, and the headline is a contradiction. The model now sits at #1 on the Artificial Analysis Intelligence Index v5.0 — ahead of Claude Opus 5.5 and GPT-6 Astra — yet as of today, developers cannot access it. The API still returns only gemini-4-argon. Carbon is rolling out to a narrow group of cybersecurity professionals first through Google’s Fairwind Program, with no public release date attached.
What Carbon Actually Does Better
Two weeks after Argon, the improvements are real and specific. On Terminal-Bench 4.0, Carbon scores 65.2% versus Argon’s 57.1% — an 8-point jump on a benchmark that measures CLI-based coding agents. FrontierSWE 2 moves from 55.0% to 61.2%. Web development gained 89 ELO points on WebDev Arena, landing Carbon at rank #4 globally (1768 ELO). Secure code generation doubles Argon’s success rate in Google’s internal testing. Additionally, unsupported assertions — hallucinations on factual claims — are roughly cut in half.
The cybersecurity numbers are where Carbon separates itself most clearly. On SEC-bench Pro, which measures vulnerability discovery, Carbon hits 88.2%. ExploitGym — validating whether those vulnerabilities are actually exploitable — reaches 54.8%, up 12.8 points from Argon. Furthermore, CWE-bench for patch generation comes in at 71.0%. Google’s Chrome team reported Carbon produces patches more than 10 times better than competitor large frontier models.
However, one caveat worth flagging: Carbon has only 9 completed benchmark entries on independent trackers compared to Argon’s 31. Some comparisons are marked “coming soon.” Moreover, on factual knowledge tests specifically, Carbon’s hallucination rate rose from Argon’s 4% to 11% — a meaningful regression on that narrow dimension despite the broader improvements. Numbers from 9to5Google’s coverage and the official Google announcement align on the headline claims.
Why Cyber Defenders Get Access First
The access restriction isn’t arbitrary. Carbon is the strongest model Google has shipped at exploiting software vulnerabilities — 54.8% on ExploitGym, which tests whether a model can validate real exploits. A model that excels at finding and exploiting vulnerabilities is a dual-use tool by definition. As a result, Google is running it through U.S. government voluntary pre-release testing and deploying it first to Fairwind Program organizations: government agencies, healthcare providers, telecommunications companies. The logic is that cyber defenders in those sectors use the exploitation capabilities for defense, not offense.
Cloud security firm Wiz confirmed Carbon leads on web penetration testing. For most developers building applications, access to ExploitGym-leading capabilities is largely irrelevant anyway. In fact, the gated rollout is frustrating, but Carbon’s defining strengths aren’t what most development workloads actually need right now.
Pricing: What to Plan For
Carbon launches at the same introductory rates as Argon: $2 per million input tokens, $10 per million output tokens, with a 95% discount on cached inputs. The important difference is what happens next. Argon’s introductory pricing ends November 30, 2026 — at which point both models move to $4/$20 per million tokens. Consequently, Carbon’s unspecified end date gives it a pricing advantage if you get access before Argon’s clock runs out.
| Model | Input (per 1M) | Output (per 1M) | Access |
|---|---|---|---|
| Gemini 4 Carbon (intro) | $2.00 | $10.00 | Fairwind only |
| Gemini 4 Argon (intro) | $2.00 | $10.00 | Public (until Nov 30) |
| Gemini 4 Argon / Carbon (post-intro) | $4.00 | $20.00 | Public (from Dec 1) |
What to Do Right Now
Stay on Argon. It remains the best publicly available Gemini model and handles general coding workloads well. Mark your calendar for November 30 — Argon pricing doubles on December 1. Meanwhile, bookmark the Carbon announcement page or set a Google AI Studio alert to catch API availability updates. When access opens, run your actual workloads before committing. For example, Carbon’s strengths are specific — vulnerability workflows, long-horizon engineering — and Claude Opus 5.5 still leads the Coding Agent Index (62 vs Carbon’s 59) for agentic development.
Google shipped Carbon two weeks after Argon. That pace reflects how competitive the frontier model market is right now. Whatever access looks like in the next few weeks, Carbon is not the last model Google ships this quarter.
