Gemini 4 Argon vs Gemini 3.1 Pro: What’s Changed?

👁 114 reads

Gemini 4 Argon replaces Gemini 3.1 Pro as Google’s most capable model, but both remain available in the Gemini app. Here’s what changed: independent benchmark results, API prices, limits and plan availability. I also ran a small test, sending both models the same three prompts in Polish through OpenRouter. At the end, I explain whether you need the Ultra plan to use Argon. In short: usually, no.

Gemini 4 Argon vs Gemini 3.1 Pro: key differences

The OpenRouter limits are from the model pages I checked on October 13: both models have a context window of 1,048,576 tokens and a maximum output of 65,536 tokens there. Argon’s higher output limit applies to the Gemini API directly.

Benchmarks: a big jump overall, a smaller one in Polish

Artificial Analysis gives Argon 52.6 points on its intelligence index. That puts it level with GPT-6 Astra (52.7) and well ahead of Gemini 3.1 Pro Preview (30). The analysis also notes that Argon rarely hallucinates (15% on the AA-Omniscience test), but has 50% accuracy because it refuses to answer more often, rather than knowing more.

In my Lab - 13 Polish-language writing tasks scored blind by two judges - the gap is smaller: Argon scored 81.8 points, compared with 77.8 for Gemini 3.1 Pro. That makes it the best Gemini in my ranking’s history, but only eighth among cloud models. Its weakest result was text naturalness (6.50 out of 10), while language correctness was its strongest (8.78). I cover the results in more detail in Gemini 4 Argon in Polish, and reviewed the older model in my Gemini 3.1 Pro review.

API pricing: cheaper through November, then it depends on the task

On paper, the difference is small. Through November 30, Argon costs $2 / $10 per million tokens, less than 3.1 Pro’s $12 output price. From December 1, Argon’s price rises to $4 / $20 and it will cost more per token.

In practice, how much the model “thinks” also matters. At the low reasoning setting on OpenRouter, Argon used no reasoning tokens, while 3.1 Pro used 4,320 across the three tasks. Those tokens account for most of the cost difference.

My test: three prompts in Polish

On the evening of October 13, 2026, I sent the same three prompts to google/gemini-4-argon and google/gemini-3-1-pro-preview through OpenRouter, with both models set to low reasoning. The prompts below are English translations of the Polish test prompts; the test itself was conducted in Polish.

1. Formal letter:A request to a tax office director to pay outstanding 2025 personal income tax in 12 instalments, including the legal basis. The tested prompt, translated from the Polish original, specified an amount of 4,800 original Polish currency units.
2. Logic problem: Three people work at different rates; one joins later and another takes a break. What time do they finish? Correct answer: 10:25.
3. Summary: Summarise a paragraph about a public transport plan in two sentences, using no more than 45 words and including the most important figures.

Formal letter: both were correct, but Argon wrote more

Both models cited Article 67a(1)(2) of Poland’s Tax Ordinance, proposed 12 equal instalments, justified the request by the taxpayer’s important interest and listed supporting documents, including an ORD-HZ statement. Argon wrote a longer justification and added an argument based on the public interest: “This will allow me to repay the liability voluntarily without the need to start costly enforcement proceedings.” It also added instructions for submitting the letter, including through Poland’s e-Tax Office. Gemini 3.1 Pro produced a shorter, equally correct letter.

Logic problem: a tie

Both models broke the work into intervals (8:00-9:00, 9:00-9:30, 9:30-10:00 and after 10:00) and arrived at 10:25. The difference was cost: 3.1 Pro used 819 reasoning tokens and took twice as long to respond.

Summary: Argon included more, but exceeded the word limit

Argon’s translated response said: “To avoid a deficit of 19 million, a single-ride ticket will rise in price to 4.60 despite a 12% drop in passenger numbers.” It also included the purchase of 24 buses for 86 million and two new routes, but its response came to 48 words against a 45-word limit. Gemini 3.1 Pro kept to 38 words, but left out the fall in passenger numbers and the new routes. It used 2,014 reasoning tokens for this short answer, making it 14 times more expensive than Argon.

My test suggests similar quality: Argon gives fuller answers, while 3.1 Pro sticks more closely to word limits. Argon was 3.3 times cheaper and almost three times faster. I saved the full responses in JSON files.

Should you pay for Ultra?

The main change for everyday users is that Argon is not locked behind Ultra. Google makes it available in the Gemini app starting with the Pro plan. However, when I checked my Pro account in Poland on the evening of October 13, Argon was not yet available. Google was rolling it out gradually, so check with the provider for availability in your country and on your account.

Pro should be enough if:

• You use Gemini for writing, emails, summaries, document analysis and questions in the app.
• You code occasionally and are happy using chat or paying for API tokens.
• You want to choose between Argon, 3.1 Pro and 3.8 Flash in one place.

Ultra may be worth it if:

• You code every day in Antigravity and want Argon there, because the Pro plan does not include Argon in Antigravity.
• You test Argon in the AI Studio or Android Studio interfaces; according to 9to5Google, this also requires Ultra.
• You regularly hit the Pro plan’s limits.

If you only want to try Argon, don’t buy a subscription in advance. My three OpenRouter prompts cost $0.02. I’ve covered the Ultra plan options in Google AI Ultra pricing, and you can find everything about the new model in my Gemini 4 guide.

// read moreGemini 4 Argon Hallucinates Sources? A 16-Paper TestBenchmarksGemini 4 Argon: How to Use the Model and APIGuidesGemini 4 Argon in the App: Rollout Status and ChecklistNews

Kinetyra