Google put another frontier model on the clock eleven days after Argon. Carbon arrives with cyber scores in the open and a developer surface that is still closed.
On 28 October 2026, Koray Kavukcuoglu announced Gemini 4 Carbon as Google's strongest model yet for cybersecurity and long-horizon work. The stake for buyers sits past the index rank: what "cyber-first staged release" means when the public API is empty and the linked safety paperwork still points at Argon.
Google posts Gemini 4 Carbon with Fairwind-first access
Google's 28 October post frames Carbon as the next frontier step after Gemini 4 Argon. The headline numbers are vendor-reported: Artificial Analysis Intelligence Index v5.0 at 59, SEC-bench Pro at 88.2% for finding vulnerabilities, ExploitGym at 54.8% for exploitation, and Collinear's CWE-bench at 71.0% for patching. Those cyber benches were run without guardrails, so production traffic will not see the same numbers.
The rollout is deliberate and gated. Google says it is in the US government's voluntary pre-release testing process. Starting the week of launch, Carbon goes to a subset of cyber defenders, then gradually to additional trusted partners through the Fairwind Program. Broader developer, business, and consumer surfaces are promised "as soon as possible," without a dated cutover.
Pricing is printed. Introductory rates match Argon at $2 per million input tokens and $10 per million output tokens, with cached input at a 95% discount on the input rate. After the introductory window, Google says $4 / $20 will apply. The end date of the intro window is not disclosed on the launch post.
Internally, Google reports Carbon running for days on multi-step engineering work, more than 2x uplift versus Argon on secure-code generation, an almost 80% drop in high-severity misalignment alerts in live coding with its own teams, Terminal-Bench 4.0 successful cheating at 0.0%, and 0.0% attack success on Gray Swan's Indirect Prompt Injection benchmark. WebDev Arena is listed at 1768 Elo, fourth place, up 89 Elo.
Fairwind partners and cyber-defence teams inherit the first seats
Google DeepMind and Google shipped the model and the public write-up. Fairwind Programme partners and the named cyber-defence cohort inherit first access. Chrome's internal patching comparison and Wiz's web penetration-testing colour travel as customer-adjacent endorsements, still without a public API SKU for ordinary buyers.
Procurement, security engineering, and evaluation teams who wanted Argon-class capability on a meter inherit a waiting room. Antigravity and other Google product surfaces that Varun Mohan has tied to Fairwind expansion inherit the same gate: Fairwind widens before wider product availability, still without a date on the launch post.
Cyber scores and a staged defender gate after Argon
Argon landed about two weeks earlier. Carbon's public delta is the combination Google stresses: #1 on Artificial Analysis v5.0 at 59 (Argon at 50 on the same index in later roundups), named SEC-bench Pro / ExploitGym / CWE-bench leads, long-horizon coding claims, Argon-matched intro pricing, and a cyber-defender-first Fairwind path rather than a same-day API drop.
Independent nine-day readouts add the commercial stake. Carbon remains off the Gemini API pricing page through early November. getaibook's 6 November catch-up cites SemiAnalysis GDPx v1.0 with Carbon at 10.9% overall and about $16.29 average cost per task against Opus 5.5 at 12.0% and about $51.41. That cost story depends on which token rate GDPx billed and on domains where rivals' guardrails hold them back; remove cybersecurity, biology, and alignment and the order reshuffles.
On AA-Omniscience, Carbon's listed score rises while its hallucination rate also rises (11% versus Argon's 4% in that readout). Index leadership and "answers more" can travel together with "invents more." Treat that as a re-eval trigger before calling Carbon a free upgrade from Argon.
No public API, no Carbon model card, guardrail-off cyber benches
Access is Fairwind and defender-gated. A public Gemini API listing, AI Studio SKU, and consumer surface remain unsettled on the launch post and on independent checks through about 6 November. A Carbon-specific model card is not linked from the launch; close readings report the linked safety material still covering Argon. Context window, output limit, and parameter count are not disclosed on the announcement.
Cyber benchmark figures are without guardrails. The Chrome "more than 10X better patches than another lab's large frontier model" claim names neither the rival nor a public harness. Whether Google has published a Frontier Safety Framework critical-capability determination for Carbon remains unsettled in the independent readout. Introductory pricing end date is not disclosed.
Treat Artificial Analysis 59 and the three cyber percentages as dated vendor-and-leaderboard claims. Treat "available to developers" as a promise without a calendar.
Keep Argon evals warm and diary the Fairwind queue
Platform leads who need Carbon's cyber or long-horizon claims should request Fairwind access and keep a private Argon-versus-Carbon harness ready, including a hallucination check rather than index rank alone. Security buyers should ask which Fairwind tier maps to their workload and whether production guardrails will be on for the benches they care about.
Finance and capacity planning should model both $2/$10 and $4/$20 until Google dates the intro window, and should treat GDPx's ~$16 figure as rate-sensitive. Product counsel should refuse customer-facing copy that quotes ungated cyber scores as production performance. Teams that only need a general chat model can ignore Carbon until an API row appears.