Gemini 4 Carbon: Access, Pricing and What Builders Should Do Now

Gemini 4 Carbon is Google's new top model, announced on October 28, 2026, but most builders can't use it yet. For now it goes only to a subset of vetted cyber defenders, then to more partners through Google's Fairwind Program. Google calls general availability for developers "coming soon" and gives no date. Access will likely be priced at $2/$10 per million input/output tokens at first and $4/$20 after that. The useful work right now is getting your evaluations and budgets ready, so you can decide quickly once the API opens.

The announcement drew more than 880 points and 700 comments on Hacker News. Below is what Google actually released, which numbers are independent, and a checklist for teams on Gemini, Claude or OpenAI today.

What Google announced

According to Google's launch post, written by Koray Kavukcuoglu of Google DeepMind, Carbon is "our strongest model with state-of-the-art cybersecurity and long horizon capabilities." It is built for long-running agent work, such as finding the root cause of bugs and patching them across large codebases. Google says some internal teams have had runs last days. Google also claims gains in coding, full-stack web development, scientific tasks and multimodal understanding.

Carbon arrives only two weeks after Gemini 4 Argon. Constellation Research reads that as a sign Google is releasing frontier models faster. (We covered the earlier model in our Gemini 4 Argon guide.) Carbon uses the same safety stack as Argon: safety training, real-time misuse defenses, prompt-injection resilience and chain-of-thought monitoring. Google says the company is also taking part in the U.S. government's voluntary pre-release model testing.

Key facts at a glance

Who can actually use it

Carbon is gated because of what it can do in security. Google says the same skills that help defenders find and fix vulnerabilities could also help attackers. So the rollout starts with a hand-picked group of defenders and then widens through the Fairwind Program.

Fairwind's own page says priority goes to "governments and organizations critical to societal resilience." That means national cyber authorities, critical infrastructure operators (healthcare, telecom, energy, finance) and core technology platforms. Academic labs doing defensive benchmarking may also apply. Google says the program has more than 650 partners worldwide. Partners must:

1. Pass background checks and a review of their ethical track record.
2. Enforce user-level authentication, phishing-resistant MFA and access controls.
3. Limit access to internal security, incident response or penetration testing teams, and track how employees use it.
4. Never share, redistribute or sell access.

Permitted uses are dual-use defensive work: authorized threat simulation, reverse engineering and malware analysis. Writing malware is prohibited. As of October 29, the Fairwind page still lists Gemini 4 Argon as the program's model and doesn't mention Carbon yet.

Realistically, if you run a typical SaaS or AI product, you won't qualify. You'll get Carbon when it reaches general availability.

How long that might take

Argon offers the only precedent. WinBuzzer reports that Argon went to Fairwind partners first, announced on September 30, and reached broader access, including the paid developer API, on October 13. Android Headlines gives October 14 for the global rollout, likely because of time zones. Argon took about two weeks from gated preview to the public API. Carbon may follow a similar path, but that's an inference from one data point, not a Google commitment.

For consumer products, the Argon rollout went to AI Pro subscribers ($20/month) in the Gemini app and Search AI Mode. AI Studio and Antigravity access was reserved for AI Ultra (from $99/month), according to Android Headlines. Google hasn't said whether Carbon will follow the same tiers.

Pricing: plan for $4/$20, not $2/$10

Google's post gives an introductory price equal to Argon's: $2 per million input tokens and $10 per million output tokens, with cached input 95% off. After the introductory period, the price becomes $4 input and $20 output. Google hasn't said when Carbon's introductory period ends. Argon's discounted rate ends on November 30, 2026, per Android Headlines and WinBuzzer. Vals AI already lists Carbon at the $4/$20 standard rate.

Two practical points:

• Carbon uses more tokens per task. WinBuzzer, citing Artificial Analysis data, reports that Carbon used about 200 million tokens across the Intelligence Index, against 110 million for Argon. At standard rates, the weighted cost works out to $7.76 per index task for Carbon versus $7.68 for Argon. The per-token price is the same, but Carbon spends more tokens per task, so its per-task cost isn't lower.
• Caching matters more for long-horizon agents. Multi-day agent runs re-send large shared prefixes. At a 95% discount, cached input costs $0.10 per million tokens at the introductory rate and $0.20 at the standard rate. Structure your prompts so the stable parts (system prompt, repository map, tool definitions) come first.

A quick worked example: an agent step with 100K input tokens and 20K output tokens costs $0.40 at introductory rates ($0.20 input + $0.20 output) and $0.80 at standard rates, before caching. Run your own traffic numbers through the LLM API cost calculator at both price points before you promise customers a margin.

Benchmarks, and which ones to trust

Google reports these results in its post:

Google also claims more than twice Argon's secure-code success rate, about half as many unsupported claims, and a drop of almost 80% in misalignment alerts in internal coding environments. Google says its Chrome team saw patch generation more than 10x better than "another lab's large frontier model."

The caveats

The independent numbers show a much closer race than "#1" suggests. Per WinBuzzer's breakdown:

• On the Artificial Analysis index, Carbon scores 59, against 58 for Claude Opus 5.5, 57 for GPT-6 Astra and 50 for Argon. Carbon ran at its high reasoning setting; Opus and Astra ran at maximum.
• On the Coding Agent Index, Carbon in Google's Antigravity CLI scores 59, behind Claude Code with Opus 5.5 at 62.
• On FrontierSWE 2, Carbon scores 61.2%, behind GPT-6 Astra (65.5%) and Opus 5.5 (62.3%). On the offline OSWorld subset, Opus 5.5 leads 87.2% to 81.5%.
• Artificial Analysis measured Carbon's hallucination rate at 11%, up from Argon's 4%, though it still calls that the second-lowest it has recorded.
• Google ran its external security evaluations without guardrails, so restricted versions may not perform as well. Google hasn't published the underlying data for the Chrome team and Wiz comparisons, or a completion rate for its multi-day tasks.

Vals AI's per-task results show the same pattern. Carbon ranks #1 on Vals' Finance Index and Legal Research Bench, but 8th of 121 on Vibe Code Bench (89.06%) and 4th of 57 on Terminal-Bench 4.0 (63.13%). Vals also lists an average latency of 47 minutes 45 seconds per Vals Index run. That's a long-horizon reasoning model, not something for interactive chat.

In short, Carbon looks strongest at security work, long autonomous runs and some finance and legal research tasks. For general coding agents and computer-use automation, it's roughly level with Opus 5.5 and GPT-6 Astra, not clearly ahead.

What this changes for builders

If you build security tooling (SAST, dependency scanning, pentest automation, incident response), Carbon is the release to watch. If your company is a core technology platform or serves critical infrastructure, apply through the Fairwind interest form now. Google gives no response timeline. If you don't qualify, expect your enterprise customers to start asking whether you support Carbon once it reaches general availability.

If you build coding agents or developer tools, don't rewrite your roadmap around a model you can't call yet. Get ready to evaluate it within days of general availability. Independent coding-agent results put Opus 5.5 slightly ahead today, so the honest answer for your users may be "we'll support it as an option," not "we're switching."

If you build long-running agents, Carbon's claimed multi-day runs, lower overreach (down 26%) and reduced benchmark cheating on Terminal-Bench 4.0 are relevant. They're also hard to verify from outside. Your own logs from long tasks are worth more than any vendor chart.

If you mainly care about cost, Carbon at $4/$20 costs the same per token as Argon but uses more tokens. For high-volume, latency-sensitive work, cheaper models like the ones in our AI coding assistants directory will often be the better fit.

What to do now

1. Build a model-agnostic eval set from 30–100 of your real tasks, with pass/fail checks. Then a new model's arrival is an afternoon of testing, not a debate.
2. Abstract the model ID. Google hasn't published Carbon's API identifier, so keep model names in config, not code.
3. Budget at $4/$20. Treat the introductory price as a bonus. Remember that Argon's discount ends November 30, 2026, if Argon is already in your stack.
4. Restructure prompts for caching so long-horizon runs benefit from the 95% cached-input discount.
5. Check the Fairwind page and Google's Gemini model announcements for general-availability news. Carbon isn't listed on the Fairwind page yet, and an update there would likely be one of the first signs of broader access.

Things to watch next: an official general-availability date and model ID, whether Carbon's introductory pricing has an end date, independent security evaluations run with guardrails, and whether Anthropic or OpenAI retake the top of the Artificial Analysis index. WinBuzzer notes that Google's previous leads there each lasted only weeks.