Gemini 4 Argon: High Potential, High Cost

Google's newly announced Gemini 4 Argon leads independent benchmarks for business workflows and takes the top spot on the Valse Index.

The model demonstrates a notable improvement in admitting when it lacks information, enhancing its reliability, but its coding results are inconsistent, and a complex pricing structure significantly alters its value comparison.

Previous Gemini releases, such as Gemini 3.1 Pro, exhibited high reasoning scores but failed to deliver expected performance in practical agent testing, particularly in coding applications.

Practical coding failures included generating duplicate methods that led to compiler errors, and Gemini 3.5 Flash was criticized for a significant price increase without proportional performance gains, alongside delays in the Gemini 3.5 Pro release, which contributed to skepticism regarding current launch claims.

The initial rollout of Gemini 4 Argon is exclusively available through Google's Fairwind program, targeting selected cybersecurity partners.

Google plans wider access for paid API customers and Google AI Ultra subscribers, but the company has not yet provided a public release date, and current subscriptions do not guarantee immediate access to the Argon model.

On Artificial Analysis, Argon achieved a high reasoning score of 53 on the Intelligence Index, matching GPT-6 Astra at its maximum reasoning capacity and exceeding GPT-6.1 Sonnet by one point.

However, Anthropic’s Opus 5.5 and Sonnet 5.5 recorded higher scores on this specific index, while the Valse Index, which prioritizes sectors like finance and law, ranked Argon at 68.9%, suggesting Argon is more economical at $15.68 per test compared to Sonnet's $21.34 and Opus's $32.14.

Argon demonstrated strong performance on Zapier's AutomationBench, achieving a 51.29% score at high reasoning, surpassing Sonnet 5.5's 44.75% with its default fallback routing.

The benchmark assesses end-to-end task completion, not merely model self-reporting, although approximately half of all tasks still require manual oversight, indicating areas for improvement.

Argon scored 38.3% on the Surge GDP.xlsx benchmark, outperforming Opus 5.5 (30.3%) and Sonnet 5.5 (29.1%).

These scores are rubric-based rather than reflecting perfect completion and, while supporting professional credibility, further testing on complex workbooks with cross-sheet references remains necessary.

In the Deep Way U 1.1 benchmark, Argon led with 77.9%, surpassing Opus 5.5 (74.2%) and Astra (74.1%), but its performance declined in Frontier's Way version 2, where it scored 55% behind Astra's 65.5% and Opus's 62.3%.

Google's own Terminal Bench 4 revealed Opus outperforming Argon with 66.4% compared to Argon's 57.4%, illustrating how methodological differences in agent setup complicate direct comparisons.

Vala’s independent tests positioned Argon second in app building and code migration, trailing only Sonnet 5.5.

Artificial Analysis reported Argon's hallucination rate at 15%, which is notably low, reflecting a higher propensity to admit ignorance rather than generate incorrect responses.

While Argon's knowledge accuracy stands at 50%, lower than Astra's 63%, this trade-off is considered beneficial for research-heavy tasks where reliability is paramount.

Argon secured the number one position on the September 30th Arena text leaderboard, achieving a score of 1525 from approximately 4,900 votes.

This human preference metric is distinct from technical performance and task-based metrics, reflecting user satisfaction in factual knowledge and code base modification.

The million-token output limit provided to Gemini 4 Argon significantly expands the model's capacity for longer reasoning trajectories, allowing for more complex processing.

Artificial Analysis's 'long decode continuation' tests, which involve pausing and resuming responses through follow-up calls, confirm this feature effectively prevents request timeouts in demanding optimization and reasoning tasks.

Argon scored 84.2 on the F1 metric in tests evaluating graph-walking capabilities across 256,000 to 1 million tokens, surpassing Astra’s score of 71.8.

This result is promising for tasks involving complex relationships across extensive inputs, though one graph benchmark cannot fully establish its understanding of all large code repositories.

Google reported a 91.7% score for Gemini on LVBench, but its testing methodology used one frame per second for Gemini while applying fixed frame caps for competitors due to API limitations.

This difference in input handling provides a practical advantage for Gemini, making it an unequal test of visual reasoning despite the high reported score.

On CWE Bench version 1, Argon matched Astra and Grok 4.7 with a 68% score in programmatic single attempts, while a separate judge panel measure showed Argon at 62% compared to Opus at 67%.

Benchmark cost estimates, ranging from $63 to 79 cents, are considered unreliable due to variations in agent harnesses and cached pricing, complicating direct cost-efficiency comparisons.

Google advertises introductory pricing for Gemini 4 Argon at $2 per million input tokens and $10 per million output tokens, with a 95% discount on cached input.

However, a footnote clarifies that standard rates will double to $4 and $20, respectively, after the promotional period, prompting advice to budget based on these future standard rates rather than the initial discounted prices.

Artificial Analysis metrics indicate Argon costs $1.99 per intelligence index task at its promotional rate, a figure projected to rise to $3.98 after the discount period ends.

In comparison, GPT-6.1 Soul is significantly more cost-effective at approximately $0.72 per task, highlighting that Argon's lower unit price is offset by its higher output volume, generating 62,000 tokens per task versus Astra's 27,000.

Despite its inconsistent coding performance, the model is a credible contender, but its adoption in workflows should be based on practical, real-world results rather than solely on benchmark scores, necessitating thorough testing for difficult, specific migration tasks.

The evidence presented establishes Google as a serious competitor in the AI landscape, overcoming earlier disappointments with this model's capabilities.