Google's Gemini 4 Carbon Tops Cursor's Coding Benchmark at Half the Price
Google's new frontier coding model tops CursorBench at 60.4% and ships in Cursor at half price through the end of November.
- Gemini 4 Carbon is now available in Cursor, topping CursorBench at 60.4% on high effort.
- Launch pricing is $2/M input and $10/M output tokens through November 30.
- Prices double to $4/M and $20/M on December 1 after the promo ends.
- Carbon edges out Fable 5.2 Max (60.0%) and Opus 5.5 Max (57.8%) on CursorBench 4.0.
- Previous Gemini 4 Argon scored 52.8% at default effort, matching Opus 5.5.
- Google claims #1 on CursorBench, WANDR, and AutomationBench at over 2X lower cost.
Cursor has added Gemini 4 Carbon to its model roster. At Cursor’s highest reasoning setting, the model scored 60.4% on CursorBench 4.0, the best result on the company’s internal coding-agent leaderboard. Google also says Carbon ranked first on CursorBench 4.0 and Perplexity’s WANDR research benchmark while costing less than half as much as the next-best model. Cursor’s published table does not list Carbon’s per-task cost, so the leaderboard alone cannot verify that cost claim.
Through November 30, Cursor is charging $2 per million input tokens and $10 per million output tokens. Those rates double to $4 and $20 on December 1.
Inside the 60.4% score
CursorBench 4.0 evaluates ambiguous, multi-file coding tasks drawn from real Cursor sessions. The benchmark targets the planning, tool use, debugging, and repository navigation expected from a coding agent rather than isolated code generation.
HumanEval primarily tests self-contained function completion, while SWE-bench measures issue resolution in selected open-source repositories. CursorBench reflects tasks and tooling from Cursor’s editor environment, which may make it relevant to Cursor users while also tying results to the company’s agent harness. It remains a first-party benchmark rather than an independent evaluation.
Cursor exposes four reasoning-effort levels: minimal, low, medium, and high. Higher settings allocate more computation to planning and problem-solving, which can increase latency and token consumption. Carbon’s 60.4% applies only to the high setting. Results for the other configurations are available on the Cursor evals page.
A narrow lead with an undisclosed task cost
Carbon’s 60.4% score leads Fable 5.2 Max by 0.4 percentage points in the published CursorBench table, while its average cost per task remains undisclosed.
| Model and configuration | Score | Cost per task |
|---|---|---|
| Gemini 4 Carbon High | 60.4% | Not published |
| Fable 5.2 Max | 60.0% | $32.79 |
| Opus 5.5 Max | 57.8% | $13.43 |
| Fable 5.2 High | 56.9% | $9.70 |
| Gemini 4 Argon High | 52.9% | $4.57 |
Cursor has not published confidence intervals or run-to-run variance with these figures, so the 0.4-point gap between Carbon and Fable 5.2 Max does not establish a decisive accuracy advantage.
Carbon improves by 7.5 points over Gemini 4 Argon High in the same table. Cursor’s earlier Argon announcement reported 52.8% at default effort and the same $2 input and $10 output rates now offered for Carbon’s promotion. Default-effort and high-effort results are not directly comparable.
The lead changes with the benchmark
Google positions Carbon as its flagship model for cybersecurity, long-running agent workflows, software engineering, scientific reasoning, and other complex knowledge tasks. External and vendor-reported evaluations produce a mixed ranking.
| Benchmark | Reporting organization | Reported results |
|---|---|---|
| Terminal-Bench 4.0 | Artificial Analysis | Carbon High: 65.2%; Fable’s best listed run: 61.6% |
| FrontierSWE v2 | Anthropic | Fable: 69.9%; Astra: 65.5%; Carbon: 61.2% |
These benchmarks use different task sets, agent harnesses, scoring methods, and reporting organizations. Their scores cannot be combined into a single ranking. Carbon’s clearest current advantage is its promotional token price, while the accuracy leader varies by evaluation.
The discount changes the math
A request using 100,000 input tokens and 20,000 output tokens costs $0.40 at the promotional rates: $0.20 for input and $0.20 for output. The same usage costs $0.80 after the December 1 increase, before any separate platform or tool charges.
Agent costs vary with context length, retries, tool calls, generated output, and reasoning effort. Token rates therefore provide only part of the comparison; cost per completed task is the more useful production metric.
- Match the effort setting. Reproducing the headline score requires Carbon High. Minimal, low, and medium are separate operating points with different latency, cost, and accuracy.
- Use identical tasks and tools. Keep prompts, repository state, permissions, time limits, and agent configuration fixed across models.
- Measure completed work. Track test pass rates, regressions, developer interventions, wall-clock time, tool calls, and total tokens per successful task.
- Model both price periods. Evaluate the promotional rates and the doubled December rates before changing a production default.
- Test representative repositories. CursorBench may resemble real editor work, but each codebase has its own languages, build systems, dependencies, and failure modes.
Build the case with your own workload
Teams using Cursor with Opus, Sonnet, or Gemini 4 Argon can run a fixed regression suite against Carbon High during the discount period. Carbon’s published score supports testing it for difficult multi-file work, but the small lead over Fable and the missing per-task cost leave the production decision to workload-specific measurements.
Comments
