Google’s published introductory API rates for Gemini 4 Argon are lower than OpenAI’s listed GPT-6 Astra rates, but coding results split by benchmark. For US developers, the prices provide a clear comparison of token charges; they don’t determine which model will cost less for every completed task. A Reddit user’s Unity-project report offers a separate, personal view of subscription usage.
Published API rates: input, cache, and output
Google’s October 13 launch post lists introductory Gemini 4 Argon rates of $2 per million input tokens and $10 per million output tokens. It also offers cached input at 95% below the introductory input rate, or $0.10 per million tokens. OpenAI’s GPT-6 Astra API documentation lists $10 per million input tokens, $1 for cached input, and $50 per million output tokens.
The published rates are in US dollars per million tokens. Google’s introductory Argon rates apply through November 30, 2026; Google says the rates will be $4 per million input tokens and $20 per million output tokens afterward.
| Model and period | Input per 1M tokens | Cached input per 1M tokens | Output per 1M tokens |
| Gemini 4 Argon, introductory period | $2 | $0.10 | $10 |
| GPT-6 Astra, listed rates | $10 | $1 | $50 |
Astra has a separate price condition for long inputs: requests above 272,000 input tokens are charged at twice the standard input and cached-input rates and 1.5 times the output rate for the full request. That can materially change the bill for a long-context job.
For a straightforward token-bill example, an uncached request with 200,000 input tokens and 50,000 output tokens comes to $0.90 at Argon’s introductory rates and $4.50 at Astra’s listed rates. Those calculations exclude tools, retries, taxes, and other charges. The unit prices help estimate API spend, but the actual bill also depends on the tokens a task uses and the applicable rates.
Coding benchmarks favor different models
The reported coding results change with the evaluation. AceCloud’s comparison lists Gemini 4 Argon at 77.9% and GPT-6 Astra at 74.1% on DeepSWE v1.1. On FrontierSWE v2, the same comparison lists Argon at 55.0% and Astra at 65.5%.
| Benchmark | Gemini 4 Argon | GPT-6 Astra |
| DeepSWE v1.1 | 77.9% | 74.1% |
| FrontierSWE v2 | 55.0% | 65.5% |
AceCloud’s benchmark comparison reports different results across these evaluations, with differing result provenance and conditions. The practical takeaway is specific: Argon leads on the reported DeepSWE v1.1 figures, while Astra leads on FrontierSWE v2. Neither row settles how the models will perform on a developer’s own codebase.
The broader index comparison is also a different measure. Artificial Analysis lists Argon at 53 on Intelligence Index v4.3.2; Intelligent Living reports the same displayed score for Astra, with Argon at High and Astra at Max reasoning. That index result sits alongside—not in place of—the task-specific coding results.
The question of how benchmark scores relate to hands-on coding also arose in our earlier coverage of concerns about Gemini 4 Argon’s practical coding performance. The results above offer a narrower comparison: two named coding evaluations, with different leaders.
One Unity project, one user’s subscription experience
Reddit user AlexXx52 reported a read-only Unity-project investigation in which Argon High received a subjective 9.8/10 after 17 minutes, while Astra High received 9.3/10 after 14 minutes. AlexXx52 also reported using about 200% of the Argon subscription allowance and about 20% of the Astra allowance. The user cautioned that the providers’ allowance percentages are not equivalent units of compute, so those figures describe that account’s experience rather than a like-for-like measure of model consumption.
