The launch chart for Artificial Analysis Coding Agent Index v2.1 shows Gemini 4 Neon (xhigh) in Antigravity CLI at 59 points, with an average cost of $6.86 per task. Claude Code with Fable 5.2 and Opus 5.5 shows 58 points for both, with average costs of $15.47 and $12.61 per task, respectively. Original X post
What this evaluation can assess: How the listed coding agent configurations perform on code changes, terminal tasks, and repository Q&A, and what they cost to evaluate.
Limits on extrapolation: Differences between models using the same agent framework, success rates in a team's own repositories, and tool subscription costs.
Model versions: Gemini 4 Neon, Gemini 4 Carbon, Claude Fable 5.2, and Claude Opus 5.5. The original post leaves the API model snapshots unspecified.
Test environment or client: Neon and Carbon use Antigravity CLI. Fable 5.2 and Opus 5.5 use Claude Code. The original post leaves the CLI versions unspecified.
Reasoning tiers and parameters: Neon and Fable 5.2 use xhigh, Carbon uses high, and Opus 5.5 uses max. The original post leaves temperature and other parameters unspecified.
The index contains 322 tasks, each run three times. Artificial Analysis averages the three scores for each task, weights tasks equally within each benchmark, and then averages the three benchmark scores with equal weights. FrontierCode-AA 2.0 has 132 code modification tasks. A submission that fails a blocking criterion earns 0 points. Other submissions are scored against the review rubric. Terminal-Bench 4.1 has 66 terminal tasks and scores them on whether the tests pass. SWE-Atlas-QnA-AA has 124 repository Q&A tasks, and each answer must meet every scoring criterion. Coding Agent Index v2.1 methodology
Cost per task is also averaged first at the task level and then at the benchmark level. The methodology page says Artificial Analysis uses each provider's API token prices, usually taking into account token usage and measurements of typical cache hit rates. Cost calculation methodology
Using the whole-number scores shown in the chart, Neon scores 43 on FrontierCode-AA 2.0, below Fable 5.2 at 49 and Opus 5.5 at 44. Neon scores 68 on the SWE-Atlas-QnA-AA repository Q&A benchmark, ahead of their scores of 62 and 66. Chart from the original post
The chart shows index scores of 59 for Neon and 55 for Carbon, a difference of 4 points between the displayed integers. The original English post and the launch analysis both say Neon leads by 3 points. Both leave the underlying unrounded scores unspecified. Original X post · Artificial Analysis launch analysis
| Agent framework | Model and reasoning tier | Coding Agent Index | FrontierCode-AA 2.0 | Terminal-Bench 4.1 | SWE-Atlas-QnA-AA | Average cost per task (USD) |
|---|---|---|---|---|---|---|
| Antigravity CLI | Gemini 4 Neon (xhigh) | 59 | 43 | 65 | 68 | 6.86 |
| Antigravity CLI | Gemini 4 Carbon (high) | 55 | 38 | 65 | 64 | Unspecified |
| Claude Code | Claude Fable 5.2 (xhigh) | 58 | 49 | 64 | 62 | 15.47 |
| Claude Code | Claude Opus 5.5 (max) | 58 | 44 | 65 | 66 | 12.61 |
The table uses the whole-number scores from the chart and the costs from the original English post. It covers only the four configurations named in the post itself. Original X post · Chart from the original post
Each row represents a combination of model, agent framework, and reasoning tier. Comparisons with Claude Code change both the framework and the model, and some also change the reasoning tier. These comparisons cannot isolate any single factor's effect on the score. At the time of the original post, Artificial Analysis said Neon in Antigravity CLI was available to Pro and Ultra subscribers. The dollar figures in the table estimate costs from API token usage per task. Subscription costs depend on the plan. Agent settings · Original X post
Use the chart from the original post to verify the launch index and the whole-number scores for each benchmark. Check the English post text for the costs of the three configurations. To rerun the evaluation, match the agent framework, model, reasoning tier, task set, and grader, and average three attempts per task as specified in v2.1. The methodology page publishes the repository Q&A data and grading configuration. The 132 FrontierCode-AA 2.0 tasks are unpublished, so a full rerun of the index requires materials beyond those publicly available. Chart from the original post · Benchmark methodology and task sets
Gemini 4 Neon