Gemini 4 Carbon: the model card

"Gemini 4 Carbon" is the query for Google's newest frontier model. Unlike most model-word pages we build — where the only verifiable source is a directory description — this one starts from a first-hand capture of Google's official announcement, "Gemini 4 Carbon: our next frontier model," published October 28, 2026 and signed by Koray Kavukcuoglu, SVP at Google DeepMind and Chief AI Architect at Google observed. Google's framing: Carbon is "our strongest model with state-of-the-art cybersecurity and long horizon capabilities" vendor claim. This page is the model card we can build from that announcement: what Google states, what is vendor-reported versus independently verified (spoiler: nothing is independently verified yet), and what the announcement never says. Labels: vendor claim / observed / analysis.

What Google announced, in its own words

• Positioning. "The next frontier model: Gemini 4 Carbon, our strongest model with state-of-the-art cybersecurity and long horizon capabilities" vendor claim.
• Family context. Carbon builds on Gemini 4 Argon, which Google launched two weeks earlier observed — a two-week cadence between flagship-family announcements analysis.
• Long-horizon engineering. Google says Carbon is "built for sustained agentic workflows," with internal teams seeing it "run for days to persist through dead ends" on multi-step tasks like root-causing and patching subtle bugs across large codebases vendor claim.
• Deliberate, phased rollout. Google says it is "actively engaged in the U.S. government's voluntary process for pre-release model testing," will make Carbon available first to a subset of cyber defenders, then phase access to additional trusted partners through its Fairwind Program, and gather feedback "before making Carbon available to developers, businesses, and consumers ... as soon as possible" vendor claim.
• Price as announced. Carbon launches at the same introductory price as Argon: $2 per million input tokens, $10 per million output tokens, with cached input at 95% off input price vendor claim. That is Google's launch statement, not a permanent price sheet — introductory prices change.

Model card (from the official announcement)

Field	What the announcement says
Name	Gemini 4 Carbon observed
Maker	Google DeepMind observed
Announced	October 28, 2026 observed
Category	Frontier model; strongest in Google's lineup vendor claim
Headline strengths	Cybersecurity (detection, validation, remediation) and long-horizon agentic work vendor claim
Predecessor	Gemini 4 Argon (launched ~2 weeks prior) observed
Announced pricing	$2/M input, $10/M output; cached input at 95% off (introductory, same as Argon) vendor claim
Availability	Phased: cyber-defender subset → Fairwind Program partners → broader surfaces vendor claim
Context window	Not stated in the announcement we captured analysis
Independent benchmarks	None exist yet — every number below is Google-reported analysis

The numbers Google reports — all vendor-side, none independent

Google published a full benchmark slate on launch day. These are Google-reported figures vendor claim — useful as claims to track, not as settled facts. No third party has replicated them yet analysis:

Benchmark (per Google)	Reported result	What Google says it measures
SEC-bench Pro	88.2% vendor claim	Finding real vulnerabilities autonomously
ExploitGym	54.8% vendor claim	Validating which vulnerabilities are exploitable
CWE-bench (Collinear's)	71.0% vendor claim	Patching at root cause with working code fixes
Artificial Analysis Intelligence Index v5.0	#1 at 59 vendor claim	Third-party index Google cites for its #1 ranking
WebDev Arena	#4 at 1768 Elo (+89) vendor claim	Web development quality in an arena setting
Terminal-Bench 4.0	Cheating 0.0%, attempts 0.4% vendor claim	Alignment behavior under evaluation pressure

Google also states internal results: more than 2x uplift over Argon in secure-code generation success rate, ~80% fewer high-severity misalignment alerts in internal live coding, and roughly 2x fewer unsupported claims in internal testing vendor claim. Internal benchmarks are the least checkable kind — we list them so you know what the vendor is claiming, not because you should cite them analysis.

What the announcement does not say

• Context window. Not in the text we captured. Details await official documentation analysis.
• General availability date. "As soon as possible" is a promise shape, not a schedule analysis.
• Independent verification of any benchmark. Launch-day numbers are vendor-reported by definition; replication takes weeks or never happens analysis.
• Consumer pricing beyond the introductory token rate. Subscription tiers and enterprise terms were not in the announcement analysis.

How to treat a two-week-old frontier model

• Separate the announcement from the access. Announced October 28 does not mean available to you; Google's own rollout is explicitly phased through the Fairwind Program first observed.
• Price the workload, not the launch post. When you get access, run your three real workloads and cost the month in our API cost calculator with the announced rates — introductory prices move analysis.
• Watch for independent replication. The #1-on-AAII claim is checkable when the index updates and others re-run it; until then it is a vendor-cited ranking analysis.
• Compare against the same-week competition honestly. OpenAI shipped GPT-6.1 Terra two days earlier; see our cross-vendor page Gemini 4 Carbon vs GPT-6.1 Terra — no head-to-head exists yet, and any page claiming a winner is ahead of the evidence.

Related pages

• Gemini 4 Carbon vs GPT-6.1 Terra — the same-week cross-vendor comparison, without invented head-to-heads
• Genie 4 — Google DeepMind's world-simulation model, announced the day before Carbon
• GPT-6.1 Terra — OpenAI's October 26, 2026 announcement
• Use cases — where frontier and specialized models earn their keep

Gemini 4 Carbon FAQ

Is Gemini 4 Carbon available to everyone?

No — Google's announcement describes a phased rollout: a subset of cyber defenders first, then trusted partners via the Fairwind Program, before developers, businesses and consumers vendor claim. Availability through Google's product surfaces is not something we can verify for your account; check Google's official pages analysis.

Is Carbon better than Gemini 4 Argon?

Google positions Carbon as its strongest model and reports domain-specific uplifts over Argon — but every delta is Google-reported vendor claim. The honest read: Carbon is the newer flagship in the same Gemini 4 generation, announced two weeks after Argon observed.

Should I build on Argon now or wait for Carbon access?

Build on what you can actually call. Argon is launched and priced; Carbon access is phased observed. Waiting on a rollout queue is not an engineering strategy — evaluate Argon on your workload and revisit when your Carbon access lands analysis.

Are the cybersecurity benchmark numbers real?

They are real claims, published by Google with named benchmarks and settings ("evaluated without guardrails," per the post) vendor claim. Whether they hold under independent replication is exactly what nobody can tell you yet analysis.

Where will verified specs appear?

On Google's own model documentation first. We will update this card with labeled, sourced facts — including the context window — when they exist analysis.

This page is descriptive and independent, built from a direct capture of Google's official announcement, last reviewed November 1, 2026. Not affiliated with Google or DeepMind. "Gemini" is a trademark of Google; benchmark figures above are vendor-reported and labeled as such.