Gemini 4 Carbon: Benchmarks, Pricing, and Access

Gemini 4 Carbon is Google's new frontier model, announced October 28, 2026: it ranks first on Artificial Analysis Intelligence Index v5.0 at 59 (as cited by Google), will launch at the same $2 input / $10 output per million token introductory price as Gemini 4 Argon, and is not yet available to developers. Access starts with a subset of cyber defenders, with more trusted partners phased in through the Fairwind Program and a wider rollout "as soon as possible" and no date.

Official Sources#

What Shipped#

Google calls Carbon "our strongest model with state-of-the-art cybersecurity and long horizon capabilities." It builds on Gemini 4 Argon, which Google says launched two weeks earlier. The announcement gives no API model id, no context window and no release date for developer surfaces, so there is nothing to put in a config file yet.

What is concrete is the rollout. Google says it is "actively engaged in the U.S. government's voluntary process for pre-release model testing," is giving Carbon to a subset of cyber defenders this week, and will phase in more trusted partners through Fairwind before opening it to developers, businesses and consumers.

Benchmarks (all vendor-reported)#

Every number below comes from Google's post and was measured by or for Google. Treat them as claims until independent runs appear.

The line that matters most for agent builders is not a score. Google says cheating on Terminal-Bench 4.0 fell to 0.0% successful and 0.4% attempted, with accuracy up, and that high-severity misalignment alerts in internal live coding dropped by almost 80%. Those are Google's internal measurements on its own teams, so they say little about your repository, but they point at the failure mode that burns long agent runs: a model that games the test instead of fixing the code.

Pricing and Context#

Google does not say when the introductory period ends. A worked example using only those published rates, and assuming the 95% cache discount applies to the later price too: an agent run that reads 2M input tokens (80% from cache) and writes 200K output tokens costs about $2.96 at $2/$10 ($0.80 uncached input, $0.16 cached, $2.00 output) and about $5.92 at $4/$20. That is the arithmetic to put in your routing model before Carbon is reachable. The cache discount carries most of the saving, which is why prompt caching across providers is worth setting up first.

What To Do Monday#

You cannot call Carbon, so do the work that makes the day it appears a one-hour job.

1. Eval Argon now. Google's developer post says Argon is available on the paid Gemini API tier in AI Studio, and Google AI Ultra subscribers can use it in Antigravity. Google says Carbon builds on Argon, so an Argon harness is the closest thing to a Carbon harness. Our earlier Argon guide was written when access was narrower; the October 13 post supersedes its availability warning.
2. Pin a private task set. Pick five tasks you can score without a model judge: a bug with a failing test, a long-context question with a known answer, a vulnerability triage with a written expected outcome. Record pass rate, tokens and retries.
3. Make the model id a config value. Carbon's id is not published. Put it behind one setting so the swap is a one-line change, as in the model routing strategies post.
4. Budget both prices. Run your token counts through the intro and post-intro rates above.

If you work in Antigravity, watch the model picker rather than the API docs; Google's Argon post says Ultra subscribers could try it there.

Which One To Pick#

• Hardest long-horizon engineering, security work: Carbon is the model Google built for it, once you can reach it. Until then, Argon at the same intro price is the testable stand-in.
• Cheap daily agent work: Not Carbon. Google's own pricing puts it in the frontier band. Compare GPT-6.1 Terra at $1/$5, which is generally available.
• Already on Claude Code: Nothing here changes your default. See Fable 5.1 as the Claude Code default and wait for independent runs of Carbon before moving.

What People Are Actually Saying#

The Hacker News thread is large and mostly about access, not capability.

• The availability complaint. One commenter argued Artificial Analysis should only rank models people can use, since Google can claim the top score for a model nobody can call. Another noted that recent Anthropic and OpenAI models reached users on or near announcement day, while Argon took weeks between announcement and availability.
• The vetting caveat. A reply pointed out that if Google is still vetting Carbon, post-training or filtering changes could shift the benchmark score before general release, as happened with other labs' staged models.
• The honesty claim. A commenter singled out the 0.0% successful-cheating result as probably the most important line in the post for coding agents. Another was struck that a lab wrote "fabrication" rather than "hallucination."
• The counter-case. A minority defended the approach: some called a staged release the standard for frontier models and said vetting who gets access is reasonable, others noted Argon did ship quickly, and one said Google simply wants to announce the top spot while it holds it. The thread was still mostly negative on the delay.

My read: the benchmarks are probably real and probably not yet decisive. What changes your next two weeks is whether Carbon reaches the public API before the next Claude or GPT release, and Google has not said.

FAQ#

What is the API model id for Gemini 4 Carbon?#

Google has not published one. The announcement says Carbon is coming to developers "as soon as possible" and does not name an id, so keep the model id as a configuration value.

How much does Gemini 4 Carbon cost?#

Google says $2 per million input tokens and $10 per million output tokens during an introductory period, rising to $4 and $20 afterwards, with cached input 95% off the input price. No end date for the introductory price is given.

Can I use Gemini 4 Carbon today?#

Not as a general developer. Google is giving it to a subset of cyber defenders and phasing in trusted partners through the Fairwind Program. Gemini 4 Argon, its predecessor, is available on the paid Gemini API tier.

Is Gemini 4 Carbon better than Claude or GPT for coding?#

Google reports state-of-the-art security benchmarks and a #4 WebDev Arena placement, all vendor-reported. No independent head-to-head on real repositories exists yet, so run your own tasks once access opens.

Sources#

Last updated: October 30, 2026

Continue Reading#

• Gemini 4 Argon: What Developers Can Actually Use - the predecessor Carbon builds on, and the model you can test today
• GPT-6.1 Terra: $1/$5 Pricing, Benchmarks, and When to Pick It - the cheaper generally available alternative
• Google Antigravity in 2026: IDE, Antigravity 2.0, CLI, Pricing and Limits Explained - where Google surfaces its newest models first
• Prompt Caching Across AI Providers: Prices, TTLs and Setup - how to capture the 95% cached-input discount
• AI Model Routing Strategies for Cost-Effective Coding in 2026 - keep the model id swappable
• Gemini - the model family, pricing, and alternatives on one page
• Claude vs Gemini - how the two model families compare for developers
• Android Studio Now Runs Claude and Codex: ACP Setup Guide - Rabbit 2 hosts 40+ ACP agents and Android CLI installs as a Claude Code and Codex plugin