Gemini 4 Argon review: Opus-level product data at a quarter of the cost, until you ask it to sell

Gemini 4 Argon is Google's first model above the Flash line in seven and a half months, and it is a big step up from Gemini 3.1 Pro on every benchmark we ran. On our product listing test it passed as many products as Claude Opus 5.5, for a quarter of the cost and in under half the time.

The trouble starts when the brief changes. Asked to write sales copy for the same products, it added five times as many unsupported claims. Its slide decks kept the numbers but none was ready to present, and its survey questionnaires needed fixing. We ran six of our benchmarks on 13 and 14 October at the API default, medium thinking. Here is what we found.

Gemini 4 Argon at a glance

• Use it for structured product data and extraction at volume, where it is close to the best models for a fraction of their cost.
• Check it for marketing copy, slide decks and survey design, where it trails Claude Opus 5.5 and GPT-6.1 Sol.
• Price: $2/$10 per million input/output tokens until 30 November, then $4/$20. Budget at the December price.
• Speed: the quickest of the top models on our listing test, but too slow at medium thinking for our faster chess clocks.

What Google says

Google announced Argon on 30 September for a small group of security partners, and opened it to everyone on 13 October: the paid Gemini API, AI Studio, Gemini Enterprise, and the Gemini app for Pro and Ultra subscribers. The model card describes a new sparse mixture-of-experts model, not a fine-tune of an earlier one, with a March 2026 knowledge cutoff.

Google's comparison table puts Argon first on 14 of 19 benchmarks against GPT-6 Astra, Claude Fable 5.1 and Claude Opus 5.5, including the Vals Index (68.9%) and long-context retrieval up to a million tokens. It does not lead on the harder coding tests: FrontierSWE v2 (55.0% against Astra's 65.5%) and Terminal-bench 4.0 (57.4% against Opus's 66.4%). Those are the developer's claims, all run at high thinking, and the table has no Gemini 3.1 Pro column.

Gemini 4 Argon benchmarks, measured by Spring Prompt

The Gemini 4 Argon model page has every result as it lands. How to read this table: each row is one benchmark, with Argon's score at the default medium setting, its rank among the models we have tested on it, and the best model on that benchmark.

Scroll sideways for all 5 columns →

Ranked on the release we publish with this post. A group only counts models that benchmark has measured. "Best" is the top model in that group, not necessarily one you would choose instead: see the alternatives. Artificial Analysis has not measured Argon's speed yet, so there is no speed group.

Product data: level with Opus, at a quarter of the cost

CatalogBench gives a model 56 invented products, with source data and images, and asks for a listing that could go live with no human fix. We run each product three times. Argon passed 70.8% of listings, the same as Claude Opus 5.5 and Grok 4.7, and was right in all three repeats for 60.7% of products, ahead of Opus's 55.4%. GPT-6.1 Sol and GPT-6 Astra reached 75.0%, and GPT-6 Sol and Muse Wonder 66.1%, which puts Argon 5th of 22.

It did that for $1.83 a run, against $6.72 for Opus and $6.24 for Astra, and in 7.3 seconds a listing, against 16.9 and 12.3. GPT-6.1 Sol was cheaper still at $1.41 and added no unsupported claims at all; Argon added about 23 per run. Gemini 3.1 Pro reached 44.6% on the same test.

Asked to sell, it made things up

The same products with a sales brief is where Argon fell back. Unsupported claims rose from 23 per run to 111, and the share of products reliably publishable dropped from 60.7% to 17.9%, 9th of 22. Opus behaves similarly (21.4%), so this is a common failure rather than Argon's alone, but GPT-6.1 Sol held at 67.9% with four unsupported claims. If Argon writes your product marketing, check every claim against the source.

Surveys and slide decks: much better than 3.1 Pro, still behind

On SurveyBench, which asks a model to plan a survey and then analyse supplied responses, Argon scored 84.5/100: 89.3 for the plan and 82.4 for the analysis. That is far above Gemini 3.1 Pro's 71.7, and behind Claude Haiku 5.5 (90.9), Opus 5.5 (96.3) and GPT-6.1 Sol (97.2). Its plans contained 20 question flaws across 12 tasks, seven of them questions that ask two things at once. When judges compared its work head to head with other models' on the same tasks, it rated 964, 5th of 10, behind Opus 5.5 (1,497), Muse Wonder (1,494), GPT-6.1 Sol (1,450) and Haiku 5.5 (1,276).

On DeckBench, its eight slide decks traced 95% of numbers back to the source data, but none was judged ready to present: six failed on layout and four used draft-looking charts. Head to head it rated 998, 13th of 21 and last of the six models with similar intelligence; GPT-6 Astra leads on 1,553, Opus 5.5 rates 1,409 and Gemini 3.1 Pro 763.

Ad budgets and chess against the clock

In ROASBench, a 12-month simulation of a skincare brand's ad budget, Argon scored 47.0/100, 8th of 21, between Gemini 3.8 Flash (51.4) and Gemini 3.1 Pro (43.5). The leaders, GPT-6 Astra and GPT-6 Sol, scored 56.6 and 54.1. This is one simulated run per model, not evidence of real advertising returns.

BulletBench plays speed chess against a ladder of engines, so slow thinking loses on time. In blitz, three minutes plus two seconds a move, Argon took a median 1.9 seconds a move and rated 1,147, the highest of any model at that clock, though over eight games the interval runs from 979 to 1,373. With one minute a game, medium thinking was too slow: it lost 7 of 8 games on time and rated 147, 19th of 23. Low thinking did not rescue it: it lost its four screening games on time at both one minute and ten seconds plus one, as did medium at ten seconds. The one-minute leader is Gemini 3.8 Flash at low thinking, at 1,068. This measures decisions against a clock, not general intelligence.

Third-party results

Every third-party figure so far uses high thinking, not the API's medium default.

• Artificial Analysis (reported 30 September, our fetch 14 October): 52.6 on the Intelligence Index, level with GPT-6 Astra and up from Gemini 3.1 Pro's 30. It used about 62,000 output tokens per task against Astra's 27,000, so its cost per task is closer to Astra's than its token price suggests, and it doubles in December. That puts it 5th of 200 models available today, behind Claude Opus 5.5 (57.6), Claude Sonnet 5.5 (56.0), Claude Fable 5.1 (53.4) and GPT-6 Astra (52.7). AA's launch analysis also reported a 15% hallucination rate, against 51% for Astra. Data sourced from Artificial Analysis.
• Arena (reported 30 September, early votes): first on the Text leaderboard with 1,525 and eighth on WebDev.

Pending: Arena's live leaderboard for the generally available model, tau2-bench, Vectara's hallucination leaderboard, and Artificial Analysis's speed measurement.

What people are saying

A selection, with how we read each one:

Gemini 4 Argon vs Claude Opus 5.5, GPT-6.1 Sol and GPT-6 Astra

These are the models the model page puts beside it, plus GPT-6.1 Sol, which costs the same blended price and leads our product listing test, and its predecessor. Prices are blended per million tokens at today's rates; Argon's rises to $8 in December. The overall score is our leaderboard, out of 100.

Scroll sideways for all 6 columns →

Claude Opus 5.5 is the better model on our business benchmarks and on the Intelligence Index, and costs twice as much today; from December the per-token prices are the same, so Opus becomes the default choice unless speed matters. GPT-6.1 Sol costs the same blended price, beat Argon on product listings, surveys and decks, and stayed factual with a sales brief. Argon's case is speed and, until December, price: it was the fastest of the top models on our listing test, and Arena voters ranked its writing first.

Is Gemini 4 Argon good? Who should use it

• Product data at volume. Extraction, attribute filling and factual listings, where it matched Opus 5.5 for a quarter of the cost and answered faster than any of the top ten.
• Interactive features that need a strong model quickly, if a few seconds is acceptable. At medium thinking it is not a sub-second model.
• Not for sales copy without a fact check. It added 111 unsupported claims per run with a sales brief. GPT-6.1 Sol added four.
• Not for finished slide decks or survey questionnaires without review. Opus 5.5 and GPT-6.1 Sol are clearly better at both.
• Budget at the December price. At $4/$20 it costs the same per token as Opus 5.5, and it uses more tokens than most models to get there.

As always, these benchmarks aren't your data, so test it on your own prompts before switching. It is in all 28 of our playground tests, so you can compare its output with other models' on the same prompt.

What we'll update

We'll publish a week-one verdict once Arena, tau2-bench and the hallucination leaderboards have results for the generally available model, and again when the price rises on 1 December. Until then, the Gemini 4 Argon model page has every result as it lands.

Choosing or switching the model behind a product feature? Our custom evaluations test the models you are weighing on your own prompts and data, against your own definition of good. If AI writes your product content, the catalogue feed diagnostic measures how far you can trust it. Talk to us, or email hello@springprompt.com.

Quick answers