Cursor adds Gemini 4 Argon with a 1M-token output limit

Cursor has added Gemini 4 Argon, making Google’s top-tier coding and reasoning model available in its editor and agent workflows. At Cursor’s default medium effort, Argon scored 52.8% on CursorBench, Cursor’s coding-agent evaluation, roughly matching Anthropic’s Opus 5.5 in the reported results.

Introductory pricing starts at $2 per million input tokens and $10 per million output tokens, about half the comparable Opus 5.5 token rate. That advantage could make Argon useful for long agent runs, though its high token consumption can narrow the effective savings.

The launch math

Cursor plans divide usage between Cursor’s models and an Other Models pool for third-party systems. Argon draws from the Other Models pool at the listed per-million-token rates. A request using one million uncached input tokens and 100,000 output tokens would cost about $3 at the introductory price.

Expected post-introductory rates are $4 per million input tokens and $20 per million output tokens. The same example would then cost about $6, bringing Argon closer to Anthropic’s pricing tier.

Effort changes latency

Cursor exposes four effort settings: minimal, low, medium and high. Medium is the default and produced the published 52.8% CursorBench result. Low and minimal prioritize response speed for straightforward edits, while high allocates more computation to difficult debugging, planning and refactoring tasks.

The visible generation rate excludes some initial reasoning time, while the end-to-end figure includes more of the full request. The published benchmark covers medium effort only, so it does not quantify how much accuracy high effort adds. Teams should compare both settings on representative repositories before accepting the additional latency and token use.

Long trajectories are the draw

Google increased the maximum output from 64,000 tokens to 1 million. That ceiling allows an agent to sustain much longer coding trajectories, reducing continuation requests during repository-wide migrations, generated test suites and multi-file refactors.

The practical output available in a request can still be constrained by the total context window, Cursor’s implementation and other product limits. Large generations also increase cost and review time, so the maximum is more relevant to autonomous agent runs than routine editing.

In a Google demonstration, Argon agents modified an existing Rust port by replacing 32,000 lines of SIMD code. The agents ran profile-guided experiments and inspected compiler output. Google reported that the resulting memory-safe decoder ran 2.7 times faster than the earlier Rust port while producing identical video output.

Benchmark wins come with caveats

Google’s published comparisons show Argon performing well on long-horizon software tasks, legal and financial work, long-context retrieval and business automation. A tally across 19 reported rows gives Argon 13 wins, one tie and five losses against GPT-6 Astra, Claude Fable 5.1 and Claude Opus 5.5. Those rows cover different evaluations, so the count should not be treated as a single aggregate score.

• Reported evaluations still place GPT-6 Astra or Opus 5.5 ahead on several tests.
• Artificial Analysis used 110 million Argon tokens to run its index, compared with a median of 82 million across models. Higher token consumption reduces the benefit of a lower per-token price.
• Artificial Analysis tested Argon at high effort while running most competitors at their maximum settings, which complicates direct comparisons.
• Axios reported that Bloomberg had heard concerns from some Google employees about internal performance. Google disputed that account.

Access widens after a narrow launch

Argon initially launched through Google’s Fairwind Program for trusted cyber defenders and internal teams. It was absent from OpenRouter, models.dev, Vertex AI, Gemini CLI, Cursor and GitHub Copilot documentation during that early window. Google said broader access would begin with paid API customers and Google AI Ultra subscribers.

Cursor’s integration gives developers a major third-party interface for using the model on real repositories. Its placement in the Other Models pool also makes usage and costs visible alongside other external providers.

Early posts on the Chinese developer forum linux.do praised the CursorBench score but questioned the price and the limited apparent difference among effort levels. Those reports are anecdotal, yet they identify two useful evaluation criteria: total tokens consumed per completed task and the measurable benefit of high effort over medium.

Choosing Argon inside Cursor

Teams evaluating Argon should track completed-task cost, first-token latency, total output tokens and test-pass rates on their own codebase. The strongest use case is sustained work across large repositories, especially when fewer continuation requests offset the model’s tendency to generate more tokens.