Google's Gemini 4 lineup clarifies the difference between Carbon and Argon

The Gemini 4 naming confusion

Users searching for 'Gemini 4 Pro' are often looking for a product that does not officially exist in Google's API model list. While Google uses positional naming for its tiers—where Pro typically denotes the general-capability flagship—there is no specific 'gemini-4-pro' string in the current catalog. Instead, the model occupying that tier is Gemini 4 Argon.

The confusion stems from how Google has historically organized its lineup. Because the 'Pro' slot has remained steady while other names like 'Flash' have shifted frequently, many people assume the next flagship will automatically carry the Pro name. This expectation has led to significant search traffic for a model that has no dedicated model card or rate card.

As OrcaRouter reported, any page claiming to offer 'Gemini 4 Pro' access is likely referring to Argon. This distinction is critical because a name without a model card behind it can lead to inaccurate comparisons. Many current online benchmarks for 'Pro' are actually just data from Argon runs or the announcement for Carbon.

Carbon vs. Argon: Performance

Gemini 4 Carbon and Gemini 4 Argon are both frontier models released within two weeks of each other in late 2026. Carbon was announced on October 28, 2026, and scored 59 on the Artificial Analysis Intelligence Index v5.0. In contrast, Argon was announced on September 30, 2026, and scored 50 on the same index.

The gap between the two models is most visible in specific knowledge work tasks. Carbon achieved an Elo gain of 126 on the AA-Briefcase benchmark, reaching a score of 1,609 compared to Argon's 1,483. This represents the largest single improvement in the launch, highlighting Carbon's strength in complex information processing.

However, Carbon also shows some trade-offs in reliability. It has an 11% hallucination rate on the AA-Omniscience v1.1 test, which is higher than Argon's 4% rate. While Carbon is often more accurate, it also tends to provide fabricated findings with more confidence, a factor that may impact security workflows.

Access and pricing details

The most significant difference for most users is how they can actually use these models today. Gemini 4 Argon is generally available on Google's paid API as of October 13. It includes a documented context window of over one million tokens shared between input and output.

Gemini 4 Carbon is currently restricted to the Fairwind Program and a specific subset of cyber defenders. It does not have a public endpoint or a published model ID. While it is built on the same architectural lineage as Argon, it is not yet available for general commercial integration by the public.

Both models share an identical introductory price list of $2 for input and $10 for output. For Argon, this introductory rate is scheduled to end on November 30, 2026, at which point the standard rate of $4 input and $20 output will apply. Carbon's pricing clock has not started because it lacks a public endpoint.

What to watch for next

There is a possibility that the distinction between these two models might blur in the coming months. Some monitors have flagged potential capability changes on the Gemini 4 Argon alias recently. If Google updates the Argon weights with Carbon's post-training, the two models might effectively merge into a single endpoint.

Because there is no technical report from Google to confirm the architecture, it remains unclear if Carbon is a completely new model or just a refresh of Argon. Some analysts suggest it looks like a slightly better version of Argon rather than a fundamental shift in how the model functions. This lack of documentation keeps the technical relationship between the two open.

For developers currently using Argon, the most important dates to track are the price increases in December and any further updates to the Argon alias. Using automatic failover across different providers is a way to manage these changes. This allows users to switch to a different model if the primary one begins behaving differently after a weight update.