Artificial Analysis records Argon (High) at 53 on Intelligence Index and 42 on AA-Omniscience. Output speed is unavailable; this snapshot does not establish general latency or all-task performance.
Suitable for judging: Comparing the overall intelligence, knowledge reliability, hallucination metrics, public pricing, and evaluation cost of a high-reasoning version within the same Artificial Analysis snapshot.
Not suitable for extrapolating: General success across all tasks, production latency, actual billing, performance on Chinese-language tasks, or direct rankings across different benchmark harnesses.
Applicable model version: Gemini 4 Argon (High); the page explicitly identifies it as a reasoning version and notes that a non-reasoning variant may exist.
Test environment or client: The Artificial Analysis Intelligence Index and its component evaluations; the page does not disclose Argon’s call endpoint or complete harness, and related charts still display “No publicly available API.” That widget does not establish current access: Google’s official catalog lists the stable API model.
Reasoning level and parameters: The page title and charts label it high; Technical specifications shows Reasoning: Yes. The current Prompt Options show Parallel Queries: Single and Prompt Length: 10k tokens. The page also offers Multiple (1k tokens only), 1k, and 100k options, but this collection did not switch options or rerun the evaluation.
Artificial Analysis describes Intelligence Index v4.3.2 as incorporating 10 evaluations: AA-Briefcase v1.1, GDPval-AA v2.1, AutomationBench-AA, Terminal-Bench 4.0, SciCode, Humanity’s Last Exam, GDP.pdf, CritPt, AA-Omniscience, and AA-LCR v1.1. The page’s charts are dynamic snapshots, and the model set differs by metric. For example, the Intelligence Index chart displays 27 of 699 models, while the AA-Omniscience chart displays 27 of 573 models.
The page publishes metric definitions, but not Argon’s per-task inputs, outputs, complete prompts, call logs, random seeds, failure handling, provider snapshot, or a unified independent reproduction package. The figures below should therefore be treated as Artificial Analysis measurements under its method and snapshot, and should not be compared directly across harnesses or dates.
| Metric | Directly visible result for Argon (High) | Page definition and limitation |
|---|---|---|
| Artificial Analysis Intelligence Index | 53; summary rank 8/216; Intelligence 4/4 | v4.3.2 composite index; the page says it is above the comparable-model median of 27; not a general score for all tasks |
| AA-Omniscience Index | 42 | Measures knowledge reliability and hallucination; ranges from −100 to 100; rewards correct answers, penalizes hallucinations, and does not penalize refusals |
| AA-Omniscience Accuracy | 50% | Proportion of correctly answered questions regardless of whether the model chooses to answer; chart set shows 27/573 |
| AA-Omniscience Hallucination Rate | 15% | Lower is better; proportion of incorrect answers among all non-correct responses, defined as incorrect / (incorrect + partial answers + not attempted) |
| Output tokens from Intelligence Index | 110M | The page says this is more verbose than the median of 85M; this is output during the evaluation, not a fixed model output limit |
| Speed | N/A | The page summary directly shows Output tokens per second as N/A; no tok/s value should be reported |
| Intelligence Index cost | $1.99 / task | The page calls this the average cost of evaluating one Intelligence Index task; it is not a fixed cost for production requests |
| Input / output price | $2.00 / 1M input tokens; $10.00 / 1M output tokens | The page also shows a Cache Discount of 95%; prices may vary by provider, date, and billing definition |
| Cost | #81/112; Cost 3/4 | The cost rank and rating are snapshots of the current dynamic leaderboard |
| Verbosity | #76/114; Verbosity 3/4 | The page presents 110M Intelligence Index output tokens for this dimension |
| Context window | 1M tokens | Technical specification; the context window should not be equated with the actual input length of each evaluation |
The page title is Gemini 4 Argon (High) Intelligence, Performance & Price Analysis. The provider is shown as Google, the model type as Proprietary model, and the release field displays only September 2026.
Technical specifications show: Reasoning Yes; input supports text and image; output supports text; Context window 1M. The page visualizes this as approximately 1,500 pages of 12-point Arial text.
The page summary says the model supports text and image input and text output, and summarizes the result with Intelligence Index 53, 110M output tokens, $1.99 per task, and $2/$10 pricing.
AA-Omniscience Index 42 cannot be interpreted as a factual accuracy rate by itself. The page defines it as rewarding correct answers, penalizing hallucinations, and not deducting for refusals.
Accuracy 50% and Hallucination Rate 15% use different denominators. Accuracy measures the proportion correct across all questions; Hallucination Rate measures, among non-correct responses, the proportion of incorrect answers when the model should have refused or admitted that it did not know. They cannot substitute for each other.
All three metrics come from the page’s AA-Omniscience chart snapshot, which lists 27/573 models. This set is not all models and is not a complete report of a general factual-verification test.
The page summary explicitly marks Speed as N/A and Output tokens per second as Unknown out of 4 units for Speed. The page also provides Intelligence Index vs. Time per Task (weighted average decode time in minutes, excluding TTFT and overhead) and End-to-End Response Time (seconds to output 500 tokens, including thinking time for reasoning models) charts, but the page text does not list a corresponding Argon value in this collection. Latency should not be inferred from chart coordinates or a missing point.
The price fields are public list prices and are not the same as the $1.99 per-task evaluation cost. The latter is Artificial Analysis’s weighted average cost for running its Intelligence Index; it cannot be directly converted into the bill for an arbitrary prompt, context length, or API provider.
Artificial Analysis’s current page supports a clear snapshot conclusion: Gemini 4 Argon (High) scores 53 on its Intelligence Index and is ranked 8/216 in the page summary; this is not a ground-truth composite capability score for all tasks.
Reliability signals must be read with their definitions: AA-Omniscience Index 42, Accuracy 50%, and Hallucination Rate 15% use different denominators and refusal handling and cannot be combined into one “accuracy rate.”
Cost has two separate definitions: $2/$10 are the public input/output prices, while $1.99 is the average Intelligence Index task cost; they are not interchangeable.
The page provides no usable Argon speed or task-time value, and Speed is N/A. Do not substitute other models’ points from the Time per Task or End-to-End Response Time charts for Argon data.
Artificial Analysis’s dynamic leaderboards, model sets, evaluation version, and model provider may change. Record the collection date, chart version, parameters, and model snapshot; do not extrapolate this page to production stability or cross-harness rankings.
For a rerun, fix Artificial Analysis Intelligence Index v4.3.2, model version Gemini 4 Argon (High), provider, Prompt Options (Parallel Queries and Prompt Length), all 10 evaluations, scoring rules, stopping conditions, failure retries, token accounting, and the price snapshot. Record the denominators and refusal handling for Accuracy, Hallucination Rate, and AA-Omniscience Index separately, and measure TTFT, decode time, end-to-end response time, and actual billing independently. If the page does not provide Argon speed or task-time data, record it as undisclosed rather than inferring it from chart coordinates.
Gemini 4 Argon
