Google now offers two Gemini 4 models: Argon and Carbon. This guide uses only what Google publishes. Prices are introductory and change after November 30, 2026, so check Google’s pages before you budget.
The two models at a glance
| Gemini 4 Argon | Gemini 4 Carbon | |
|---|---|---|
| Released to developers | October 13, 2026 | November 10, 2026 |
| Price per million tokens (Google’s list) | $2 input / $10 output | $2 / $10 until Nov 30, 2026, then $4 / $20 |
| Google positions it for | “the majority of software engineering and knowledge work” | The hardest tasks: complex multi-step coding, open-ended research, cybersecurity |
| Token use | Lower | Higher: Google says Carbon uses more tokens to reason and verify |
| Consumer access | Gemini app for Google AI Pro and Ultra | Rolling out to Google AI Ultra in the Gemini app, AI Mode and Antigravity |
A simple way to choose
- Start with Argon. Google itself calls it the right pick for most engineering and knowledge work, and it uses fewer tokens.
- Move to Carbon when a task is long and multi-step and a wrong answer is costly: a multi-package migration, a root-cause investigation, a security review of code you own, or a research question with no clear path.
- Watch the token bill. Carbon’s price per token matches Argon until November 30, but it spends more tokens per task, so the same job can cost more. From December the list price also doubles.
- Test on your own work. Google’s benchmarks are its own. Run the same prompt through both models and compare quality and cost. Our prompt workflow guide shows how.
A prompt to test both (our own)
You are reviewing a small web app I own. Here is the repository summary and the failing test output.
1. Explain the most likely root cause in three sentences.
2. Propose the smallest fix and show the changed lines.
3. List the checks you would run to confirm the fix works.
4. Say what you are unsure about.
Stop after the plan; do not rewrite unrelated files.Score each answer on whether the root cause is right, whether the fix is minimal, and how many tokens it used. If Argon gets it right, you have your default.
Briefing Carbon for long tasks (our own pattern)
Long-horizon models do best with a clear brief. Give it:
- Goal: one sentence on what done looks like.
- Limits: files or systems it must not touch, and a budget for turns or time.
- Checks: the tests or rendered pages it should verify before it stops.
- Report: what to hand back (summary, diff, open questions).
For agents that act on your systems, add the safeguards in our agent permissions checklist.
How this compares with Claude
Anthropic uses the same idea: Opus, Sonnet and Haiku 5.5 for most work and Claude Fable 5.2 for the hardest jobs. For a broader view see Claude vs Gemini and the model tracker.