We benchmarked Gemini 4 Argon against Gemini 3.1 Pro Preview on browser tasks

We'd recommend a switch to Gemini 4 Argon if your browser agent currently runs on Gemini 3.1 Pro Preview. In our Gemini 4 Argon vs Gemini 3.1 Pro benchmark experiments, each model ran 15 browser tasks, a small but challenging test set. And within the 15 attempts, Argon passed 11 of them while 3.1 Pro only passed 5.

Argon passed every single task that 3.1 Pro succeeds in, and gets there with about half the work. It used 23.9 tool calls per task when 3.1 Pro used 52.1. On long tasks across multiple sites, 3.1 Pro ran out of its 100-step budget three times without producing a result and Argon never did.

On cost, the two models are level on the same work, but Argon is cheaper once you count failed attempts as well. On the 5 tasks both models passed, Argon cost $0.588 per task and 3.1 Pro cost $0.598. Counting every attempt, Argon spent $1.62 per passed attempt and 3.1 Pro spent $4.02. However, Google's listed price for Argon doubles on December 1, 2026, so watch out for the price rise.

(Costs use Google's list prices through November 30, 2026)

Task success

On our test set, Gemini 4 Argon is the more reliable model. Argon passed 6 tasks that 3.1 Pro failed, while 3.1 Pro didn't pass any task that Argon failed. For the rest of the tasks, these two models had the same results. The 6 that Argon was better at include multi-site data collection, social media verification and search-and-compare jobs.

The two models failed in different ways. Three of 3.1 Pro's 10 failed attempts was due to running out of the 100-step budget without an answer. They all involved working on jobs across 9 or more sites. Five were simply hallucination or giving wrong answers. On a task involving inflation briefing, 3.1 Pro reported 324.000 as the 2025 annual CPI average, but that figure is actually the BLS table's second-half average. We’d be especially cautious about this kind of error: the number exists in the source, which makes it look credible, but it doesn't answer the question.

Argon passed the same task in ego(lite). And it never ran out of steps on these 15 tasks. In three of its 4 failure, it gave two values for one field, which counts as a wrong answer. During a task of information look-up regarding open-weight models, Argon gave the correct parameter total of 668.95 billion, but then it went on to offer 664.95 billion as an alternative. This extra step resulted in its failing the task. If your code reads the agent's answer, make sure that each field only holds a single value.

Cost

Through November, the two models cost about the same. But price rises for Argon when December comes. On the 5 tasks they both passed, Argon averaged $0.588 per task and 3.1 Pro averaged $0.598. Priced at 3.1 Pro's rates, Argon's usage on those tasks would have cost $0.646, so Argon's lower per-token price offsets slightly higher usage. Uncached input is $0.421 of Argon's $0.588 and $0.212 of 3.1 Pro's $0.598.

The bigger cost difference is in failed attempts. Each of 3.1 Pro's 3 attempts that ran out of steps cost $3.48 on average, 4.3 times its other attempts, and none produced an answer. Those 3 attempts were 51.8% of everything 3.1 Pro spent. Counting every attempt, Argon spent $1.62 per passed attempt and 3.1 Pro spent $4.02.

Google's Gemini API pricing page lists Argon from December 1, 2026 at $4 per million input tokens, $20 per million output tokens and $0.20 per million cached tokens. That is double its earlier price. 3.1 Pro has no announced price change.

December prices apply Google's new list price to the token usage we recorded.

From December, the extra money for Argon buys reliability. If your agent's jobs are ones 3.1 Pro already passes, 3.1 Pro will cost less per task: at December prices it was cheaper on all 5 tasks both models passed. If your jobs include ones 3.1 Pro fails, Argon passes more of them and still costs less per passed attempt, $3.25 against $4.02.

How each model works

Gemini 4 Argon often opens several sites in one browser command, while 3.1 Pro opens one page at a time and keeps going. A browser command is one script the model sends to ego-browser. Across the 15 tasks, 22.4% of Argon's 246 browser commands opened two or more pages, against 2.5% of 3.1 Pro's 489. Argon needed 16.4 browser commands per task and 3.1 Pro needed 32.6.

The central bank task shows what that means on a long job. It asks for the first rate decision of 2026 at each of 12 central banks: the date, the decision, the rate, any dissent and the next meeting date. Argon finished in 16 browser commands, and 14 of them opened between 2 and 7 banks' pages. It passed with 29 tool calls and cost $1.24. 3.1 Pro sent 73 browser commands, and none opened more than one page. It used all 100 steps, stopped while opening the 9th bank, the Bank of Korea, and gave no answer. That attempt cost $3.36.

Both models used the same browser on the same sites, so the difference came from how much each reply covered. That is what kept Argon inside the 100-step budget on jobs that walk through a list of sites.

More about Gemini 4 Argon

To explore how Argon handles more difficult jobs, we designed 13 seperate tests specifically for Argon. On a total of 28 browser use tasks, success rate drops as the number of sites involved increses.

On jobs across nine or more sites, Argon passed 8 of 14. Its highest count was 67 tool calls of the 100 allowed, never once running out of budget. To complete these tasks that involve more than nine sites, Argon cost $2.377 per task on average. Four of the six failures were wrong answers, and 2 of them gave two values for one field, confirming the previous pattern of giving extra answers.

On shorter tasks across one to three sites, Argon passed 8 of 11 at an average cost of $0.469. Two of its 3 failures gave two values for one field, and the third was a wrong answer in a web app task.

Our advice is to plan for the size of your tasks. A short lookup with Argon costs under fifty cents. A long multi-site job costs a few dollars, can run close to an hour and passed 57% of the time.

Which to use

Switch to Gemini 4 Argon if the current success rate of your 3.1 Pro is unsatisfactory. On cost, it matches 3.1 Pro per task through November and stays cheaper per passed attempt even after the December price change as it gets the job done without wasting tokens.

Staying on Gemini 3.1 Pro Preview makes sense in one case. From December 1, it costs less per task on work it already passes, and in our runs that was 5 of the 15 tasks.

We ran Argon as gemini-4-argon at a high thinking level in the pi agent harness, with ego-browser as its browser. ego-browser comes with ego(lite) and drives it as the agent's browser.

FAQ

Is Gemini 4 Argon better than Gemini 3.1 Pro for browser agents? Yes, Gemini 4 Argon passed 11 of 15 browser task attempts and Gemini 3.1 Pro Preview passed 5. Argon passed every task 3.1 Pro passed, plus 6 that 3.1 Pro failed. The gains were largest on jobs across 9 or more sites, where 3.1 Pro ran out of its 100-step budget three times.

Should I switch from Gemini 3.1 Pro to Gemini 4 Argon? Switch if your agent needs to finish more tasks: Gemini 4 Argon passed 11 of 15 attempts and Gemini 3.1 Pro Preview passed 5. Through November 30, 2026, both cost about $0.59 to $0.60 per task on work both passed. From December 1, Argon's listed price doubles, so 3.1 Pro costs less on tasks it already passes.

Is Gemini 4 Argon cheaper than Gemini 3.1 Pro? At prices through November 30, 2026, Gemini 4 Argon cost $0.588 per task and Gemini 3.1 Pro Preview cost $0.598 on the 5 tasks both passed. Counting failed attempts, Argon spent $1.62 per passed attempt and 3.1 Pro spent $4.02. 3.1 Pro's failures were expensive, because its 3 attempts that ran out of steps cost $3.48 each on average.

How much does Gemini 4 Argon cost after December 1, 2026? Google lists Gemini 4 Argon at $4 per million input tokens, $20 per million output tokens and $0.20 per million cached tokens from December 1, 2026. At those prices, our runs would cost $1.176 per task on the 5 tasks both models passed, against $0.598 for Gemini 3.1 Pro Preview. Per passed attempt, Argon would cost $3.25 and 3.1 Pro $4.02.

Does Gemini 4 Argon use fewer steps than Gemini 3.1 Pro? Yes, Gemini 4 Argon used 23.9 tool calls per task and Gemini 3.1 Pro Preview used 52.1 across the same 15 tasks. Argon opened two or more pages in 22.4% of its browser commands, against 2.5% for 3.1 Pro. Its longest task used 67 tool calls of a 100-step budget.

Why does Gemini 3.1 Pro run out of steps on browser tasks? Gemini 3.1 Pro Preview ran out of its 100-step budget on 3 of 15 tasks, all of them jobs across 9 or more sites. It opened one page per browser command, so a 12-bank task needed more replies than the budget allowed. On the same task in the same browser, Gemini 4 Argon finished in 29 tool calls.

How does Gemini 4 Argon handle multi-site data collection? Gemini 4 Argon passed 8 of the 14 tasks it ran that span nine or more sites. It never ran out of steps on them, but 2 attempts reached the one-hour time limit without an answer. These jobs averaged $2.377 per task.

What does Gemini 4 Argon still get wrong? Gemini 4 Argon failed 4 of its 15 shared attempts, and 3 of those gave two values for one field. On one task it reported the correct 668.95 billion parameters, then offered 664.95 billion as an alternative. Gemini 3.1 Pro Preview did the same in 2 of its failures, so check that each field in an answer holds one value.

Is Gemini 3.1 Pro still worth using? Gemini 3.1 Pro Preview passed 5 of 15 attempts, and from December 1, 2026 it costs less than Gemini 4 Argon on those 5 tasks. At December prices, it was cheaper on all 5. For work it doesn't already pass, Argon passed more tasks and spent less per passed attempt.

Can Gemini 4 Argon run a browser agent with ego-browser? Yes, we ran Gemini 4 Argon at a high thinking level in the pi agent harness with ego-browser, and it passed 11 of 15 attempts. The model wrote ego-browser scripts through the agent's bash tool and read the page text each script returned. Its longest task used 67 tool calls of a 100-step budget.

How we tested

We compared Gemini 4 Argon (google/gemini-4-argon) and Gemini 3.1 Pro Preview (gemini-3.1-pro-preview), both at a high thinking level. Both ran in the pi agent harness, version 1.5.0, with the same ego-browser build in ego(lite) on macOS. The model receives the ego-browser skill file, writes browser scripts through its bash tool and reads the page text each script returns. Argon was called through ZenMux and 3.1 Pro through Google's Gemini API. Argon ran on November 4, 2026, and 3.1 Pro ran on November 5, 2026.

Each model ran 28 tasks from our efficiency task set, and 15 of those tasks were run by both. The comparison uses the 15 shared tasks, with one attempt per model on each task. An attempt is one run of one task by one model. The job size section uses all 28 of Argon's tasks. The shared tasks fall into five groups, and each task is in only one group. The groups are collecting the same fields from many official sites, checking X posts against official pages, lookups on one to three sites, open-ended research and a business web app.

Each attempt had a budget of 100 steps and one hour. A step is one model reply, which usually makes one tool call. A tool call is one command the model sends through the agent. A browser command is a tool call that runs an ego-browser script, and we count a command as opening two or more pages from the page navigations in its script. An attempt passes only when its final answer meets every item on the task's answer key, and an answer that gives two values for one field fails.

Cost is each attempt's token usage priced at Google's Gemini API list prices through November 30, 2026. For Argon, those are $2 per million input tokens, $10 per million output tokens and $0.10 per million cached tokens. For 3.1 Pro Preview, they are $2, $12 and $0.20 up to 200,000 tokens per prompt, and $4, $18 and $0.40 above that. Cost per task averages the 5 tasks both models passed. Cost per passed attempt divides the cost of all 15 attempts by the attempts that passed.

Full results

Each cell is one attempt's result and its cost at list prices through November 30, 2026. "Out of steps" marks an attempt that used the 100-step budget without an answer, and "out of time" marks one that reached the one-hour limit without an answer.

Gemini 4 Argon's other 13 tasks, used in the job size section. Each cell is one attempt.

Update log: first published on November 9, 2026. Test runs on November 4 and 5, 2026.