Gemini 4 Carbon arrived just two weeks after Gemini 4 Argon opened to the public. Same generation, same price, but a clear step up. Here is what changed, test by test, and which one you should use right now.
The short version
| Gemini 4 Carbon | Gemini 4 Argon | |
|---|---|---|
| Announced | Oct 28, 2026 | Sept 30, 2026 |
| Public API | Not yet | Yes, since Oct 13 (gemini-4-argon) |
| Google calls it | "Our strongest model" | Its flagship reasoning model |
| Price (intro) | $2 in / $10 out per 1M | $2 in / $10 out per 1M |
| Price (standard) | $4 in / $20 out per 1M | $4 in / $20 out per 1M |
| Context window | Not published (tested to 1M) | 1,048,576 tokens |
| Benchmarks (Google's table) | Ahead on all 17 | Behind on all 17 |
| Who can use it | Selected cyber defenders | Paid API users, AI Pro and Ultra |
Use Argon today, switch to Carbon when it opens. They cost the same, so there is no reason to stay on Argon once you have Carbon access, and no reason to wait for Carbon before you start building.
Carbon vs Argon, test by test
| Benchmark | Carbon | Argon | Change |
|---|---|---|---|
| AA Intelligence Index v5.0 | 59 | 50 | +9 |
| Vals Index v2.2 | 70.4% | 69.2% | +1.2 |
| SWE-sweep | 9.6% | 4.3% | +5.3 |
| FrontierSWE 2 | 61.2% | 55.0% | +6.2 |
| Terminal-Bench 4.0 (AA) | 65.2% | 57.1% | +8.1 |
| Agents' Last Exam v1.1 | 77.4% | 69.0% | +8.4 |
| OSWorld-2.1 (offline) | 81.5% | 74.7% | +6.8 |
| AutomationBench | 53.9% | 51.3% | +2.6 |
| Vending-Bench 2 | $16,484 | $13,718 | +$2,766 |
| Chartography | 72.5% | 71.6% | +0.9 |
| Blueprint-Bench 2 | 58.6% | 54.4% | +4.2 |
| GraphWalks (256K to 1M) | 96.2% | 84.2% | +12.0 |
| RiemannBench | 80.0% | 76.0% | +4.0 |
| Terminal-Bench Science 0.1 | 64.8% | 52.9% | +11.9 |
| CWE-bench | 71.0% | 68.0% | +3.0 |
| ExploitGym | 54.8% | 42.0% | +12.8 |
| SEC-bench Pro | 88.2% | 83.4% | +4.8 |
Changes are in percentage points. From Google's announcement table; see our full benchmarks page for GPT-6 Astra and Claude.
The biggest improvements
- Exploiting vulnerabilities: ExploitGym up 12.8 points, the largest jump in the table. This is why Carbon goes to defenders first.
- Very long documents: GraphWalks up 12 points, to 96.2% on prompts of 256K to 1M tokens.
- Science work: Terminal-Bench Science up 11.9 points.
- Agentic coding: SWE-sweep more than doubled, and Terminal-Bench 4.0 rose 8.1 points.
- Prompt injection: on Gray Swan's test, successful attacks fell from 0.7% for Argon to 0.0%.
The smallest change is in reading charts (Chartography, +0.9), where Argon was already strong.
It behaves better, too
Google says it used feedback from Argon users to improve coding, agent work, science, full-stack web development and multimodal understanding. It also reports changes you'd notice day to day, measured in its own testing against Argon:
Google also says Carbon is better at asking for clarification when a task is ambiguous and at pointing out conflicting information before it acts, and that serious misalignment alerts in its live coding tests dropped by almost 80%.
What stayed the same
- The price. $2 / $10 per million tokens at launch, $4 / $20 after.
- The safety system. Carbon uses the same layered safety and security stack as Argon.
- The rollout pattern. Cyber defenders first, everyone else after.
- What it makes. Both write text and code. Neither generates images or video.
Which should you use?
Today: Gemini 4 Argon. It is the only one of the two you can actually use, through the Gemini API, Google AI Studio, Vertex AI or the Gemini app on a Pro or Ultra plan. Our sister site gemini4argon.com covers it in depth.
As soon as you get access: Gemini 4 Carbon. Same price, better on every published test, and easier to steer. If you keep the model name in one setting, the switch takes a minute.
Neither one makes videos.
For video and images from Google, use Veo 3.1, Gemini Omni and Nano Banana 2. They're all on Lumeta today, with one set of credits.
Carbon vs Argon questions
Is Gemini 4 Carbon better than Gemini 4 Argon?
Yes. In Google's table Carbon beats Argon on all 17 benchmarks, by up to 12.8 points, at the same price.
Does Gemini 4 Carbon replace Gemini 4 Argon?
Google hasn't said Argon is going away. For now Argon is the publicly available model and Carbon is limited to cyber defenders.
Is Gemini 4 Carbon more expensive than Argon?
No. Both cost $2 per million input tokens and $10 per million output tokens at launch, and $4 and $20 after the introductory period.
Sources: Google's Carbon announcement, Google's Argon availability post, Gemini API models, 9to5Google.