← Back to feed

Artificial Analysis published its evaluation of Gemini 4 Argon on September 30, the same day Google announced the model’s phased release through the Fairwind Program. The headline finding: Argon equals GPT-6 Astra on the Artificial Analysis Intelligence Index while costing approximately 60% as much per task at introductory pricing.

The AA Intelligence Index is a composite quality benchmark that aggregates performance across reasoning, coding, knowledge, and instruction-following tasks. It is the primary metric AA uses to rank frontier models across providers.

What the Cost Figure Reflects

Argon launched at $2 per million input tokens and $10 per million output tokens as an introductory price, rising to $4/$20 once the introductory period ends. GPT-6 Astra is priced at $10/$50 per million tokens. On a per-token basis, Argon at introductory rates is substantially cheaper than Astra. The 60% cost-per-task figure is AA’s actual measurement rather than a raw pricing ratio, reflecting that per-task token consumption differs between the two models in production workloads.

AA’s methodology measures cost per task directly from benchmark runs rather than projecting from published token rates. The gap between the per-token pricing difference and the 60% per-task figure suggests Argon uses meaningfully more tokens to complete equivalent tasks, consistent with its 1M output token limit enabling longer reasoning trajectories.

Top Three Positioning

The AA analysis describes Argon’s result as returning Google to the top three labs on the Intelligence Index. The previous top three, following the releases of Claude Fable 5.1 and GPT-6.1 Sol in late September 2026, had placed Anthropic and OpenAI above Google’s prior generation.

Argon’s DeepSWE v1.1 score of 77.9%, reported in Google’s launch materials, sits above GPT-6.1 Sol at 75.2% (high effort) and GPT-6 Astra’s range of 74.1% to 74.8% on the same benchmark. The AA Intelligence Index result aligns with that ordering.

Access Constraints

The cost-performance comparison is currently hypothetical for most practitioners. Argon is not available to standard API customers as of the date of the AA analysis. Access remains limited to cybersecurity researchers and trusted testers through the Fairwind Program. Google has described a phased expansion to paid API customers and Google AI Ultra subscribers without providing a timeline.

AA’s cost analysis is therefore forward-looking: it describes the economic position Argon will occupy when it enters general availability, not what any customer can currently purchase.