Gemini 4 Carbon vs Gemini 4 Argon: what Google's new model adds

The short answer: Argon to build on, Carbon to plan for

Choose Gemini 4 Argon for anything you need to run now: it is in the Gemini API paid tier with a 1,048,576-token context and in the Gemini app for Google AI Pro and Ultra subscribers. Plan to move security and long engineering work to Gemini 4 Carbon when Google opens it, since Google reports clear gains there at the same introductory price.

Scope and evidence

Task: deciding whether to keep building on Argon or wait for Carbon. Argon facts come from the Gemini API model and pricing pages; Carbon facts and both models' benchmark results come from Google's Carbon announcement and the table on Google DeepMind's Gemini page, read on October 30, 2026.

Both sets of scores are Google's own reporting. Carbon cannot yet be tested publicly, and AppFor.ai ran neither model for this comparison.

Compare the same criteria

Each line contrasts what Google has published for the two Gemini 4 models and draws the practical conclusion.

• Availability. Carbon: selected cyber defenders, then trusted partners in the Fairwind Program. Argon: Gemini API paid tier, Google AI Studio and the Gemini app. What this means: Argon is the only one you can deploy.
• API price per million tokens. Carbon: $2 input and $10 output at launch, $4 and $20 later. Argon: $2 and $10 through November 30, 2026, then $4 and $20. What this means: identical pricing, with Argon's end date known.
• Context and output. Carbon: not published. Argon: 1,048,576-token context and long decode up to 1M output tokens. What this means: only Argon's limits can be planned around.
• Intelligence Index v5.0 in Google's table. Carbon: 59. Argon: 50. What this means: a nine-point lead on Google's headline measure.
• Security in Google's table. Carbon: SEC-bench Pro 88.2%, ExploitGym 54.8%. Argon: 83.4% and 42.0%. What this means: Carbon's gain is largest at exploit validation.
• Agentic coding and long context. Carbon: Terminal-Bench 4.0 65.2%, GraphWalks 96.2%. Argon: 57.1% and 84.2%. What this means: Google reports gains on long tasks and long inputs too.

View image

Which fits your real work?

A startup shipping a code-review agent this quarter. Argon fits because it is available with documented limits and the same price Carbon will have.

A security vendor validating exploitability of reported flaws. Carbon fits on Google's ExploitGym result, through the Fairwind Program.

A Google AI Pro subscriber working in the Gemini app. Argon fits, because Google has not put Carbon in any consumer plan yet.

Cost of getting the work done

API: 2 million input and 200,000 output tokens cost $6 on either model at the introductory rate and $12 at the later rate. Argon's introductory rate ends on November 30, 2026; Google has not given an end date for Carbon's. Argon is paid tier only, with every thinking token billed.

Tokens per task decide the real cost: Google says Carbon cuts unsupported claims and overreach, which can mean fewer retries, but it has published no per-task cost.

Try both on the same small task

Suggested check: run a fixed task on Argon now and keep the inputs, so you can rerun it on Carbon on the day it opens.

• The cause matches what you already know about the bug.
• The regression test fails before and passes after the patch.
• The model lists missing information rather than inventing it.
• Thinking and output tokens are recorded for each run.

Decision checklist

Match the lines to your plans. Specs are in our Gemini 4 Argon profile and our Google Gemini profile.

• If you must ship this month, use Argon.
• If you do vulnerability triage, request Carbon through the Fairwind Program.
• If you budget for 2027, both models are announced at $4 and $20 after their introductory periods.
• If you need documented context limits, only Argon has them so far.
• If you use the Gemini app, Argon is what you get today.

Go a little further

Related reading

• Gemini 4 Argon profile →
• Google Gemini profile →
• Gemini 4 Carbon profile →