Gemini 4 Carbon Is Gated to Cyber Defenders: Seven Moves to Make on Argon Before It Opens

Gemini 4 Carbon scored 59 on version 5.0 of the Artificial Analysis Intelligence Index, one point ahead of Claude Opus 5.5 and two ahead of GPT-6 Astra, according to VentureBeat’s launch coverage. Google announced it on October 28, two weeks after Gemini 4 Argon opened up. Unless you work in cyber defense and hold a Fairwind Program seat, you cannot call it today, and Google has given no date for when you can.

That makes the useful question narrower than “is Carbon better.” It is: what do you do in the weeks between a model topping the leaderboard and a model you can actually put in production? There are two numbers in the launch data that change the answer, and a price change on December 1 that matters more to your budget than Carbon does right now.

What Google shipped, and who can touch it

Google’s announcement calls Carbon its strongest model, built around cybersecurity and long-horizon work. Access goes first to a subset of cyber defenders, then to more trusted partners through Fairwind, while Google gathers feedback and refines guardrails before a broader release to developers, businesses and consumers. Google also says it is taking part in the U.S. government’s voluntary prerelease testing. 9to5Google reports the company’s line that Carbon is “coming soon” to everyone else, with no date attached.

The pricing is announced even though the endpoint is not: $2 per million input tokens and $10 per million output tokens during an introductory period, with cached input 95% off, rising to $4 and $20 afterward. That matches Argon exactly. Google has not said when Carbon’s introductory period ends. Argon’s does: the Gemini API pricing page shows $2 and $10 through November 30, 2026, then $4 and $20 from December 1, with cached input going from $0.10 to $0.20.

The cyber numbers are the headline Google chose. Carbon posts 88.2% on SEC-bench Pro for vulnerability finding, 54.8% on ExploitGym and 71.0% on CWE-bench for patching. Google notes those evaluations ran without guardrails, so the model a Fairwind customer gets may not score the same. Outside cyber, the gains are real but not a sweep. VentureBeat’s table has Carbon on top in 15 of 17 rows, with two notable exceptions: Opus 5.5 leads OSWorld 2.1 offline at 87.2% to Carbon’s 81.5%, and GPT-6 Astra leads FrontierSWE 2 at 65.5% to 61.2%. On Artificial Analysis’s Coding Agent Index, Claude Code running Opus 5.5 scores 62 against Carbon’s 59 in the Antigravity CLI.

So if your heaviest workload is computer use or agentic coding, Carbon’s arrival does not move you even once it opens. The routing you already have for those jobs is still the routing the independent data supports.

Two numbers the launch chart does not lead with

More right answers, and more confident wrong ones

Google’s blog says Carbon cuts unsupported claims by about half compared with Argon in internal testing, and that it flags missing requirements instead of fabricating evidence. That is a company claim about its own internal tests.

Artificial Analysis measured something adjacent and got a different shape. On its AA-Omniscience leaderboard, Carbon leads the index at 54, but its hallucination rate is 11% where Argon’s is 4%. VentureBeat reports Carbon’s accuracy on the same test rose to 58% from Argon’s 48%.

Both can be true, and the definition explains why. Artificial Analysis defines hallucination rate as “the proportion of incorrect answers out of all non-correct responses.” A model that does not know an answer can either decline or guess. Argon mostly declined. Carbon answers more of the questions it used to skip; more of those come out right, which lifts accuracy, and more come out wrong, which lifts the hallucination rate. The index rewards the trade on balance. Your workload might not.

If you run retrieval over contracts, compliance questions or anything where “I can’t find that” is a correct and valued output, a model that guesses more often is a regression even when its accuracy is higher. If you run research assistance where a human checks every answer anyway, Carbon’s trade is a straight win. That is a decision about your error costs, not about the leaderboard, and it is one you can test on Argon today because the comparison point is Argon.

It thinks longer before it says anything

The Artificial Analysis model page for Carbon lists 97.48 seconds to first token and 200 million output tokens used across the Intelligence Index. The Argon page lists 44.96 seconds and 110 million. The listed cost per index task barely moved, $7.76 for Carbon against $7.68 for Argon, so the bill per task looks similar on that test suite. The wait does not.

Nearly a minute and a half before the first token is fine for a batch job that runs overnight, and Google’s own pitch (internal teams watching Carbon work on engineering tasks for days) is a batch pitch. It is a poor fit for anything with a person waiting at a screen. Lower reasoning effort will cut that latency, and our guide to reasoning effort settings covers how to find the level where your quality holds. But the high setting is the one that produced the 59.

A plan for the weeks before Carbon opens

None of this argues for ignoring Carbon. It argues for doing the work now on the model you can reach, so the swap is a one-line change when the gate lifts.

1. Run Argon on your own tasks this week. Same family, same introductory price, a published endpoint. Fifty to two hundred real tasks from your workload, scored the way you score production. Our approach to building evals applies directly: the 59 is a hypothesis about your work, not a measurement of it.

2. Score declines separately from errors. Most homegrown evals count an answer as right or wrong. Add a third bucket for “declined or flagged missing information,” and decide in advance whether that bucket counts as success. Without it, you cannot see the shift the Omniscience data shows, and you will not catch it when you move to Carbon.

3. Model your bill at December prices, not today’s. Argon doubles on December 1 across every tier: standard goes from $2 and $10 to $4 and $20, batch and flex from $1 and $5 to $2 and $10. VentureBeat’s worked example, 10 million input and 1 million output tokens, comes to $30 at the introductory rate and $60 after. Budget the $60. Carbon’s sticker is identical, so if you price on Argon’s post-December rate, you have also priced Carbon. Then decide which jobs justify that rate and which should drop a tier; the flagship repricing of the last few months means the comparison set has moved since you last ran it.

4. Lean on the cache before the price doubles. Cached input stays at 95% off on both models, $0.20 per million after December 1. Implicit caching only, per Google’s pricing page, so the work is in prompt structure: stable system prompts and shared context at the front, variable content at the end. If your hit rate is low, raising the cache hit rate is worth more than any model choice this quarter.

5. Set a latency budget per workload now. Write down, for each job, the longest time to first token it can tolerate. Anything under a minute rules out Carbon at high effort on today’s measurements. You want that list before the release, not after a product team has already wired Carbon into a chat interface.

6. If you are a cyber defender, apply. Fairwind is the access path for both Argon’s early rollout and Carbon’s. Our Argon access and pricing breakdown covers what the program asked of applicants then. Expect Carbon’s terms to be at least as strict, given the exploit scores.

7. Keep your computer-use and coding-agent routing where it is. Opus 5.5 still leads OSWorld 2.1, and the Opus-backed Claude Code harness still leads the coding agent index. For workloads that sit between those and plain text, an opaque versus transparent model router is a better place to spend engineering time this month than waiting on Carbon.

When Google opens Carbon, the work you did on Argon is the work you rerun: same task set, same three buckets, same latency budgets, new model ID. If Carbon wins on your numbers, you switch with evidence. If it guesses more where you needed it to decline, you will see that in the declined bucket on day one instead of in a customer complaint in month two.

The leaderboard moved today. Your production stack moves when your own evals say so, and the useful part of today’s news is that you can build those evals now on a model with the same price, the same API shape and a public endpoint.