AIIQ Gemini 4 Argon scoring, benchmark coverage, and methodological boundaries

One-sentence takeaway

The AI IQ page gives Gemini 4 Argon an estimated overall IQ of 136, ranking it #7 among 164 public models. The score comes from AIIQ's own seven-dimension scoring framework, but Argon has results for only 14 of 39 tracked benchmark inputs across 6 of the 7 dimensions; missing items feed into estimates, so the result is not a complete measured score or a real-world task success rate.

Use cases

• Tasks this can help assess: Comparing overall rankings, dimension distributions, and covered benchmark results within the same AI IQ snapshot.

• Tasks this cannot be generalized to: Treating the IQ score as human IQ, treating dimension estimates as raw percentages, or inferring performance on all real-world tasks, speed, cost, or input-context capability.

• Model version: Gemini 4 Argon; the page also records a public release date of 2026-09-30.

• Test environment or client: AI IQ public model profile; the page labels the modality as text. The provider, client, and model tier used for the underlying benchmarks are not stated.

• Reasoning tier and parameters: Not stated.

Evaluation method

AI IQ places Argon into seven equally weighted dimensions: abstract reasoning, mathematical reasoning, academic reasoning, programming reasoning, computer use, reliability, and knowledge work. The page says its raw benchmark table includes only results supported by sources, then combines them into an IQ score using a shared scoring method; where direct coverage is missing, it uses conservative estimates. Dimension estimates may include statistical adjustments or priors from previous-generation models, and are not raw benchmark percentages.

The page discloses results for 14 of 39 tracked inputs for Argon, spread across 6 dimensions; missing results are not treated as zero scores. The page both summarizes externally sourced benchmark results and provides AIIQ's own dimension mapping, equal-weight aggregation, and ranking. The sections below therefore distinguish “raw benchmark values” from “AI IQ-derived values.”

Key results

• Overall IQ: 136; IQ rank: #7 / 164 public models.

• Dimension estimates: abstract reasoning 112 (0/5 covered), mathematical reasoning 141 (1/5), academic reasoning 137 (3/6), programming reasoning 153 (3/6), computer use 148 (1/5), reliability 131 (2/6), and knowledge work 131 (4/6).

• The page says programming reasoning (153) and computer use (148) are the highest among dimensions with direct coverage; both are still dimension scores that include estimates.

Raw data

The page lists the following benchmark results for Argon. It does not disclose a unified test prompt, sample size, number of runs, reasoning tier, or complete harness for these rows, so they cannot be used to reconstruct a set of independent item-level experiments.

Conclusions and limitations

1. The score of 136 and the dimension scores are AI IQ-derived ranking metrics, not Google's official capability scores or percentages from a single benchmark.

2. The page includes benchmark names such as AA and Arena.ai. It presents a sourced aggregation of results, but does not explain on this page who ran each result, which model endpoint was used, or whether Argon-specific configuration was used. The entire table therefore cannot be called an AIIQ-built test.

3. Coverage is only 14/39, and dimension estimates may use statistical adjustments or priors from previous-generation models. The ranking and high-scoring dimensions should be treated as a snapshot from that date, not generalized to uncovered tasks.

4. The page does not provide a valid cost or speed measurement for Argon. It also explicitly notes that one million tokens is an output limit claimed by Google, not a verified input context window. Such publisher-provided information does not replace independent measurement.

5. Comparisons with scores from other websites must check benchmark versions, evaluation harnesses, coverage, and estimation rules at the same time; figures from different harnesses cannot be compared directly.

Reproduction notes

On 2026-10-14, access the page above and record the overall IQ, rank, dimension coverage, and 14 benchmark rows visible at that time. The model profile alone cannot reproduce each raw result: the page does not publish item-level prompts, samples, run parameters, endpoints, randomness controls, or the complete scoring scripts. To review the results, first verify the source benchmarks, versions, and harnesses separately from AIIQ's seven-dimension equal-weight aggregation and its rules for estimating missing items.