Gemini 4 Argon (High): Intelligence, Performance and Price Analysis

What it claims

Artificial Analysis rates Google's Gemini 4 Argon (High), released September 30, 2026, as a leading model in intelligence and reasonably priced for its tier. Beyond the index score and list prices, the page gives three figures that matter for budgeting:

• Cost per task: the full Intelligence Index evaluation cost an average of $1.99 per task.
• Blended price: $1.47 per 1M tokens, assuming a 7:2:1 ratio of cache hits to input to output.
• Verbosity: it generated 110M output tokens against a peer median of 81M, about 36% more.

Its per-token prices of $2.00 input and $10.00 output sit exactly at the peer median. Any cost difference against similarly priced models will therefore come from token volume, not rate. On this evaluation it spends more tokens than its peers.

Evidence

The score of 53 against a median of 26 depends on the peer group. The methodology section says:

• Reasoning models are compared against both reasoning and non-reasoning models.
• Proprietary models are compared only within their price band, here ">$1 per 1M tokens". The band is set by a blended 3:1 input/output price.

So the median of 26 includes non-reasoning models in that price band. It is not a reasoning-only baseline. The FAQ instead describes the peer group as "reasoning models in a similar price tier". The page therefore gives two different descriptions of the baseline.

The page lists many sub-evaluations:

• Agentic coding and terminal use
• Kubernetes incident root-cause analysis
• Legal agentic work
• Medical long-context reasoning
• Spreadsheet quantitative analysis
• Hallucination rate (AA-Omniscience)
• AA-Briefcase Elo
• A Finance & Accounting Index

The fetched text does not include the scores for any of these. They exist only in the page's charts.

How to use it

• The model is proprietary. Weights are not available and Google has not disclosed the parameter count.
• It is available through only one API provider, Google's API, so there is no provider choice on price or latency.
• It takes text and image input, produces text output, and has a 1.0M-token context window.
• It is a reasoning model, so expect thinking-token overhead in output billing.

Limits

• The $1.47 blended price depends on 70% of tokens being cache hits. Workloads with little caching will cost noticeably more per token.
• The page describes the same median price as both "moderately priced" and "better than average". Read it as "at median".
• Speed metrics are named in headings, but no throughput or latency figures appear in the text.
• The model was released two days before this capture, so the figures may be revised.