Gemini 4 Carbon and Gemini 4 Argon are both reasoning models in Google's Gemini 4 family. Carbon is the frontier tier, announced with higher scores across cybersecurity, coding and safety evaluations. Argon is the one you can use on Morphic today. This page compares what Google has published about both, side by side.
Gemini 4 Carbon vs Argon: what the benchmarks say
Carbon's published improvements over Argon center on three areas: cybersecurity, calibration and safety.
Cybersecurity. Carbon scores 88.2% on SEC-bench Pro, 54.8% on ExploitGym and 71.0% on CWE-bench. In an internal secure-code evaluation, it achieved more than 2x the success-rate uplift over Argon.
Calibration. Unsupported claims are cut roughly 2x compared with Argon. Overreach is down 26% and under-service is down 29%, which means Carbon is better at knowing what it knows and when to say so.
Safety. Carbon scored 0% successful cheating on Terminal-Bench 4.0 and 0% prompt-injection attack success in published evaluations. It shares Argon's safety stack and adds additional alignment training.
Coding. Carbon is #4 on WebDev Arena at 1768 Elo, an 89-point gain over its predecessor on the same benchmark.
Intelligence. Carbon scored 59 on the AI Intelligence Index v5.0, the highest of any model at the time of announcement.
Specs side by side
| Carbon | Argon | |
|---|---|---|
| Generation | Gemini 4, frontier tier | Gemini 4 |
| Type | Reasoning | Reasoning |
| AI Intelligence Index v5.0 | 59 (#1) | Not published on this benchmark |
| SEC-bench Pro | 88.2% | Not published on this benchmark |
| WebDev Arena | #4, 1768 Elo | Not published on this benchmark |
| Calibration vs Argon | ~2x fewer unsupported claims | Baseline |
| Overreach vs Argon | Down 26% | Baseline |
| Under-service vs Argon | Down 29% | Baseline |
| Secure-code uplift vs Argon | More than 2x | Baseline |
| API pricing (intro) | $2 / $10 per M tokens | Priced through Morphic plans |
| Context window | Likely 1M tokens | 1M tokens |
| Inputs | Likely text, images, audio, video, code | Text, images, audio, video, PDF |
| On Morphic | Not yet | Copilot, Basic plan and up |
| Access | Fairwind Program (cyber defenders first) | Available now |
Argon benchmark numbers for the Carbon-specific evaluations above have not been published by Google on those same benchmarks, so the table shows "Not published" rather than a figure.
What stays the same
Both models are reasoning models that write text, not media. Neither generates images, video or audio on its own. On Morphic, Argon works through Copilot and MCP, calling the platform's image, video and audio models to do the rendering. Carbon, when available, would fill the same role, with the published improvements in security, calibration and safety.
Both share the Gemini 4 safety stack, with Carbon adding an extra round of alignment training on top.
Using Gemini 4 Argon on Morphic today
Argon runs in Morphic Copilot on every paid plan, from Basic up. Pick it in the model picker, describe the project, and it plans the steps and calls the generation models. It also connects over MCP from apps such as Gemini CLI.
For a full walkthrough of both routes, see the Gemini 4 Argon guide. For more on the model, see the Gemini 4 Argon model page.
FAQs
On the published benchmarks, Carbon scores higher than Argon in cybersecurity, coding, calibration and safety evaluations. Whether that translates to a given task depends on the task, but the direction is consistent across every metric Google has published.
Not yet. Carbon is available first to cyber defenders through the Fairwind Program. Gemini 4 Argon, the same family, runs on Morphic today in Copilot and over MCP.
Carbon's introductory API rate is $2 per million input tokens and $10 per million output tokens. Argon on Morphic is priced through Morphic plans, not per-token API billing, so a direct price comparison depends on the plan and the usage.
Google has not said. Carbon is the frontier tier, not a replacement announcement. Both are in the Gemini 4 family and serve different access and pricing points.
No. Like Argon, Carbon is a reasoning model that writes text. On Morphic, Argon calls image, video and audio models to render; Carbon would do the same once available.
The Fairwind Program is how Google is rolling out early Carbon access, starting with cyber defenders. Broader access is expected in the coming weeks.