Gemini 4 Argon Lands in Cursor and Hits Top 5 on Agent Arena: 3 Numbers That Matter

Google’s newest model is already being tested by the people who matter most to it: coders. Cursor says Gemini 4 Argon is now available in its editor, and the Arena leaderboard team says it ranks in the top 5 for Agent Arena.

Two posts on X give us three useful numbers: a CursorBench score, a price, and a leaderboard placement. Here is what they say, and what they leave out.

More on this topic: TechnoSports Gemini 4 Argon coverage.

Gemini 4 Argon: Key Details

What Cursor Announced

Cursor posted that Argon is available in the product and scores 52.8% on CursorBench at default effort. Cursor calls that on par with Opus 5.5. CursorBench is Cursor’s own benchmark, so the number shows how the model behaves inside its editor rather than across the whole industry.

The listed price of $2 per million input tokens and $10 per million output tokens makes it a mid-priced option for agentic coding. Real cost depends on how many tokens an agent run uses, which the post does not say. The post had passed 193,000 views at the time we read it.

What Arena Reported

The Arena account congratulated the Google DeepMind team and said Gemini 4 Argon (High) ranks in the top 5 for Agent Arena, pointing to a thread for the details. A top 5 placement means other models still sit above it on that board.

Leaderboard positions can shift quickly as more votes arrive, so treat the rank as a snapshot from October 13.

Why It Matters for Developers

Coding agents are the hot battleground in AI right now, and a model that matches a rival on a vendor benchmark at a clear price point gives teams something to test. If you are not a developer, the same model also reaches Gemini app users on paid plans, as explained in our Gemini 4 Argon coverage.

Our advice is simple: run Argon on your own repository before switching, because benchmark wins do not always carry over to your codebase.

What Is Not Confirmed

• Independent benchmark results outside Cursor and Arena.
• Full Agent Arena methodology and the exact rank number.
• Latency and token usage of Argon in long agent runs.
• Whether Cursor’s price is passed through at cost or has a markup.

Quick Answers

What did Argon score on CursorBench?

52.8% at default effort, per Cursor.

How much does it cost?

$2 per million input and $10 per million output tokens, per Cursor.

Is it number one on Agent Arena?

No. The post says top 5.

Source: Cursor on X, Arena on X, Google Gemini on Wikipedia, Cursor on Wikipedia.

How to Read These Benchmark Claims

Vendor benchmarks are useful signals, but they are not independent proof. CursorBench is run by the company that sells the editor, so it measures tasks Cursor cares about. A score of 52.8% against a rival at a similar level tells you the models are close, not that one is better for your work.

Agent Arena is a different kind of test. It is a public leaderboard where users compare model behaviour in agent tasks, so rankings reflect human preference as much as raw accuracy. Seeing the same model do well on both a vendor benchmark and a public board is a stronger sign than either alone.

For teams, the practical step is a short pilot. Pick five real tickets from your repository, run them through Argon and your current model, and compare how many changes you would actually merge. That is the same idea behind the merge-based evaluations that have become popular this year.

Pricing also deserves a second look. Agent runs can burn many tokens through repeated tool calls, so the cheaper per-token model is not always the cheaper per-task model. Track cost per merged change, not cost per million tokens.