Gemini 4 Argon: Vals AI Professional-Task Benchmarks and Evaluation Methodology

One-sentence takeaway

Vals records Index 68.90% ± 0.97 and Finance Agent v2 65.40% ± 0.32 in the October 14 snapshot. Ranking denominators differ; results across tasks cannot be combined into one leaderboard.

Use cases

• Suitable tasks for evaluation: The professional Agent, software engineering, legal, tax, and composite Vals Index tasks listed on the Vals AI page; observing the scores, rankings, cost per test, latency, and hyperparameters published on the model page.

• Tasks not suitable for extrapolation: General production success rates, direct rankings across different Agent harnesses, capabilities on undisclosed task sets, independent results for the official Google API, or treating dynamic rankings as fixed historical facts.

• Applicable model version: Gemini 4 Argon. The page shows a publication date of 2026-09-30 but does not provide an API model ID or verifiable snapshot string.

• Test environment or client: The benchmark suite on the Vals AI model page. The page shows Google as the default provider and warns that some benchmarks may use different providers and parameters. The complete harness for each specialist benchmark is not public on this page.

• Reasoning tier and parameters: The page's Hyperparameter Settings show provider Google, temperature 1, top-p Default, top-k Default, max output tokens 262,144, and reasoning effort HIGH. These are settings displayed on the model page, not complete run logs for every benchmark.

Evaluation method

Dynamic page and time basis

• This note records the model-page results observed on 2026-10-14 as a dynamic snapshot.

• The page shows a model publication date of 2026-09-30; its Updates entry is also marked Sep 30, 2026.

• The current model-page summary, benchmark table, and update text show different ranking denominators. This note preserves the visible readings separately, does not infer the reason for the difference, and does not combine them into one ranking.

Publicly shown common settings

The model page's expanded “View Hyperparameter Settings” section shows:

• Default Provider: Google

• Temperature: 1

• Top P: Default

• Top K: Default

• Max Output Tokens: 262,144

• Reasoning Effort: HIGH

The same page also shows a context window of 1M, a maximum output of 262K, and input/output token prices of $4.00/$20.00. The page warns: “Some benchmarks may use different provider and parameters. Please refer to the benchmark page for more information.” This note only studies the specified model page and does not extend these model-page settings into complete configurations for each benchmark.

Metrics shown on the page

For each benchmark, the model page displays fields such as accuracy, error range (for some projects), ranking, cost per test, or latency. The update text describes Vals Index as a model-suite evaluation and gives the “cost per test” for some tasks. This model page does not disclose the task list, tools, scorer, number of repeated runs, sample selection, or error-range calculation method for each specialist benchmark.

Key results

The first set of figures in the tables below comes from the currently visible benchmark table on the model page. The “Update-text reading” column records the denominator or wording in the 2026-09-30 update entry on the same page; the two are not treated as the same leaderboard at the same point in time.

Vals Index and professional tasks

Legal and tax

Software and engineering tasks

Cost and latency

• The model-page overview shows a Vals Index cost per test of $15.68 and latency of 46 min 33 sec.

• The update text also lists $57.82 per test for Code Migration, $44.90 per test for Terminal-Bench Science, $30.45 per test for SRE Bench, and $193.78 per test for CUA-bench.

• These costs use the benchmark-run methodology represented on the Vals page. The page does not provide complete token, tool-call, retry, or billing details for each benchmark, so they cannot be treated as the cost of an arbitrary API request.

Raw data

• The model-page developer field is Google, and the weight status is PRIVATE. The model page shows a 1M context window, 262K maximum output, and $4.00/$20.00 input/output prices.

• The Vals Index model-page reading is 68.90% ± 0.97, 1/47; Finance Agent (v2) is 65.40% ± 0.32, 1/78. Both readings and their error ranges are directly visible.

• The model page also lists Vals RSI Index 30.55%, 4/24, CyberBench v1.1 77.86% ± 5.30, 3/47, EMB 75.24% ± 2.46, 4/74, ProofBench v1.1 99.00% ± 1.00, 6/50, and IOI 100.00% ± 0.00, 1/44, among other results. This note expands only the projects directly related to the Vals Index, professional, legal, tax, and software-task categories in the tables above.

• The update text ranks Vals Index at 1/41 and Finance Agent v2 at 1/73, and gives the corresponding 2026-09-30 update methodology. These denominators differ from the current model-page table.

• The page's Hyperparameter Settings directly show: Google provider, temperature 1, top-p/top-k Default, max output tokens 262,144, and reasoning effort HIGH.

Conclusions and limitations

1. Snapshot-level judgment only: As collected on 2026-10-14, the model page directly verifies a Vals Index score of 68.90% ± 0.97, 1/47 and Finance Agent (v2) at 65.40% ± 0.32, 1/78; these are not permanent rankings.

2. Two visible methodologies on the page: The 2026-09-30 update text uses denominators such as Vals Index 1/41 and Finance Agent 1/73. This note does not guess the reason for the difference or calculate the two sets of rankings in one table.

3. No ranking across harnesses: The page warns that different benchmarks may use different providers and parameters. Without the complete harness for each specialist benchmark, Finance, Legal, Tax, and software results cannot be combined into a single capability ranking.

4. Settings are shown but the method is incomplete: Google provider, temperature 1, HIGH reasoning, and 262,144 max output are visible. The provider, tools, system prompt, task list, scorer, repeat count, and failure handling for each benchmark are still not public on this page.

5. Costs must be interpreted by task: The Vals Index cost of $15.68 per test and the higher cost of long-horizon tasks use the page's methodology. They do not replace the price of a single API call and cannot be directly compared for cost efficiency across different benchmarks.

6. Do not fill in unseen data: The page does not disclose an API model ID, model snapshot, item-level outputs, complete task samples, the error-range calculation method, or unified randomization and stopping criteria for the evaluations.

Reproduction notes

For a reproducible or comparable test, at minimum fix:

1. Collection time and page version; explicitly record whether the current model-page table or the 2026-09-30 update text is being used.

2. Gemini 4 Argon's exact API snapshot, provider, temperature, top-p, top-k, reasoning effort, max output tokens, and context limits.

3. The task-set version, samples and denominator, input materials, tools, system prompt, Agent harness, scorer, and failure/retry rules for each benchmark.

4. The number of runs, error-range calculation method, per-test token and tool costs, latency, and human intervention for each test.

5. Do not combine results from different providers, parameters, or harnesses into one overall score. Before comparing models, confirm that the task, version, scoring, and runtime environment match.

This page is sufficient to recheck the visible snapshot and parameters above, but not to independently reconstruct the complete experiments for the specialist benchmarks.