Argon is the one you can call[Gemini 4]
Google shipped two frontier models a few weeks apart. Only one of them has a public API.
On 13 October Google made Gemini 4 Argon available to developers and enterprises on the paid Gemini API, to Ultra subscribers in Antigravity, and to Ultra and Pro subscribers in the Gemini app. The introductory price is $2 per million input tokens and $10 per million output tokens until 30 November 2026, then $4 and $20. Cached input is 95% off. The context window is 1 million tokens, input and output together. A separate Fairwind program offers Argon without cyber guardrails. That one is an application, not the public API.
On 28 October Google announced Gemini 4 Carbon. It is not on that API. Google says access starts this week with a subset of cyber defenders, then widens through Fairwind, and only later reaches developers, businesses and consumers. The introductory price is the same $2 and $10. The Carbon note does not say when that discount ends.
The score people will quote is from Artificial Analysis. On 27 October they replaced the Intelligence Index. Version 5 adds two benchmarks, upgrades three and drops three, and 60% of it now sits on private tests or held-out answers. That day Claude Opus 5.5 (max) led at 58, and Gemini 4 Argon (high) stood at 50. A month earlier, on the previous index, the same Argon setting had scored 53. Those two Argon numbers are not the same ruler.
On 28 October Artificial Analysis scored Carbon at 59 with high reasoning and called it the new first place, 9 points above Argon. Carbon was still limited to selected users when they measured it.
- Carbon, high 59
- Opus 5.5, max 58
- Argon, high 50
The same Carbon note is easy to misread if you stop at 59. On AA-Omniscience v1.1, Carbon’s hallucination rate is 11%, up from 4% for Argon, while its accuracy rises 10 points to 58%. The higher index score is not the more cautious model. At the standard $4 and $20 price, Artificial Analysis puts Carbon at $7.76 per index task and Argon at $7.68.
Google’s Carbon note also lists cybersecurity scores, measured without guardrails: 88.2% on SEC-bench Pro, 54.8% on ExploitGym, and 71.0% on CWE-bench. Those are Google’s claims. I am not charting Terminal-Bench. The 27 October index note has Kimi K3.1 Pro first there at 67%. The 28 October Carbon note has Carbon second at 64%, behind Opus, and does not restate Kimi. Two sentences, a day apart, not one chart.
I would call Argon. It is the model with a key. Carbon’s 59 is a selected-user score on a ruler that changed this week, and on the knowledge test published beside it, Carbon guesses more often than Argon.