Gemini 4 Carbon Is #1 on AA Index — Still Fairwind-Only: Cyber Benches, $2/$10 $4/$20, and What to Do Before the API

Attributed compile (not original research)

Primary source: Introducing Gemini 4 Carbon by Koray Kavukcuoglu (SVP, Google DeepMind / Chief AI Architect, Google), published 2026-10-28

Secondary roundups: Agentic Weekly — model releases (Carbon brief, first seen 2026-10-28); ToolSift daily LLM news 2026-10-28 (Artificial Analysis Index v5.0 context — Opus 5.5 at 58 before Carbon’s claimed 59; Index scores from different days are not a clean head-to-head)

Docs check: Gemini API models (fetched 2026-10-29) lists Gemini 4 endpoint gemini-4-argon only — no public Carbon model ID

This post restates Google’s Carbon packaging in my own words, with a builder checklist for the same Fairwind-first trap as Argon (~two weeks earlier). I did not run Gemini 4 Carbon — it is not publicly callable. Every bench, price, and access claim below is vendor-reported (or clearly attributed to a named secondary source). Offline preflight: docs scrape + list-price arithmetic only.

---

Google announced Gemini 4 Carbon on October 28, 2026 — roughly two weeks after Gemini 4 Argon — as its strongest frontier model, with the headline pitched at cybersecurity defense and long-horizon engineering. The capability story is loud: Artificial Analysis Intelligence Index v5.0 at 59 (Google claims #1), cyber benches without guardrails, and internal anecdotes about multi-day agent runs. The access story is quieter and more important for builders this week: Carbon starts with a subset of cyber defenders, then more trusted partners through the Fairwind Program, then developers / businesses / consumers “as soon as possible.” There is still no public API model ID.

If you shipped a plan that assumed “Carbon drops → swap the string in models.list,” pause. Same Fairwind-first pattern as Argon. Bookmark the blog; do not block a release on it.

What’s new vs Argon (all vendor-labeled)

Google frames Carbon as building on Argon with sharper cyber and long-horizon behavior, plus tighter alignment. Treat every number as Google’s claim unless you re-run it yourself.

Intelligence Index and coding/web

Cyber defense lifecycle (evaluated without guardrails, vendor)

Internal / partner anecdotes Google cites (not your SLA): >2× uplift in secure-code success vs Argon; Chrome team claimed >10× better high-quality patches vs “another lab’s” frontier model; Wiz cited for a web pen-test lead tied to Scan for Good. Useful as direction, useless as a compliance ticket.

Long-horizon behavior (vendor narrative)

Google says internal teams saw Carbon run for days, persisting through dead ends on multi-step engineering (root-cause + patch across large / low-level codebases). That is a product promise about agent stamina, not a public max_duration parameter you can set today.

Safety / alignment deltas Google reports

Built on Argon’s multi-layered stack (training, misuse defenses, prompt-injection resilience, chain-of-thought monitoring). Carbon-specific vendor figures:

Still vendor benches and internal alerts. Your agents need your own IPI and exfiltration tests.

Access trap: Fairwind / cyber defenders first

Google states it is in the U.S. government’s voluntary pre-release testing process. Rollout order in the announce:

1. This week: subset of cyber defenders
2. Then: additional trusted partners via Fairwind
3. Later: developers, businesses, consumers on product surfaces “as soon as possible”

There is no GA date and no public endpoint string. On 2026-10-29 the public Gemini API models page still lists only gemini-4-argon under Gemini 4. Guessing gemini-4-carbon is not a migration plan.

Same builder constraint as the Argon post: do not make Carbon a production default until you have (a) a documented model ID, (b) billing/access on your project, and (c) a golden eval that beats your current callable default on your failures.

X chatter will treat “#1 on AA Index” as already shippable. Chatter is not an API key.

Pricing trap: intro $2/$10 → post-intro $4/$20 (+ cache)

Google says Carbon launches at the same introductory price as Argon:

Intro matches the familiar mid-frontier coding band. Post-intro doubles both columns. Cache at 95% off input is the lever that decides whether long agent loops stay affordable.

List-price arithmetic (my assumptions — not a bill)

Offline preflight on this box used Google’s stated rates only. Assumptions: one trajectory = 80k input + 20k output; “80% cache hit” means 80% of input tokens bill at the cached rate (5% of input list); no tax, no retries, no tool-overhead tokens.

Post-intro is an exact 2.00× on these list rates when cache assumptions match. Build board slides for both columns. Do not bake “Carbon is $2/$10 forever” into a forecast. Measure real cache hit rates before you trust the middle rows.

What to do this week (callable models still win)

Carbon’s announce does not retire Argon, GPT-6.1 Terra/Sol, Claude 5.5, or anything else you can already call.

1. Keep shipping on models with a public ID and a key that works today.
2. Prep long-horizon + secure-code evals now: multi-file root-cause, flaky repros, vuln→patch trajectories with human review gates. When a Carbon ID appears, you want 20–50 frozen tasks ready — not a scramble.
3. Watch the official models directory and release notes for a real Carbon endpoint. No ID → no production default. Do not invent one.
4. Do not block a customer release on Fairwind unless your org is actually in that program.
5. Price the post-intro column ($4/$20) in any Carbon ROI slide; assume cache misses until measured.
6. Security posture: if you ever receive a no-guardrail / cyber-defender build, isolate it. Fairwind-style access is defense capability plus staged safeguards — not “same sandbox as the public chatbot.”
7. Re-check AA Index yourself when Artificial Analysis publishes Carbon’s row; private-test resets make cross-day screenshots unreliable.

Offline check (this box, 2026-10-29 America/Toronto)

Saved under /workspace/reputation/runs/29-gemini-carbon/:

• No GEMINI_API_KEY / GOOGLE_API_KEY (or siblings) in the environment
• google.generativeai / google.genai SDKs not installed
• Public docs fetch: gemini-4-argon present; no gemini-4-carbon (or sibling) model id string
• No live models.list / generateContent call — nothing to call for Carbon publicly
• Cost table above = arithmetic on vendor list prices under stated assumptions

When Carbon might become a default (once the API lands)

Only after a callable model ID and billing access:

1. Workloads are long-horizon or secure-code heavy and you are hitting today’s failure modes.
2. A golden eval shows Carbon beating your current default on your tickets — not SEC-bench screenshots alone.
3. Cost model survives post-intro $4/$20 and realistic cache hit rates.
4. Prompt-injection / tool-exfiltration tests pass under your threat model (Google claims Gray Swan IPI 0.0% ASR — verify on your agents).
5. You can rollback to Argon / Terra / Opus / Sonnet without rewriting orchestration.

Until then, Carbon is a roadmap signal with a loud cyber story and the same Fairwind gate as Argon. Treat the blog as planning input; treat callable models as this week’s shipping surface.

---

YongBo Yu · Toronto AI engineer · https://yongbo-yu.vercel.app · GitHub YongBoYu1