In Arena's published snapshot, Gemini 4 Neon (High) ranked second with 1,336 points based on about 4,200 agent sessions from real users. Its median cost per task was $2.82, 53% lower than third-place Claude Opus 5.5 (Max), which trailed Neon by one point. Arena's original post · Leaderboard chart
What this evaluation can assess: Compare user feedback on task outcomes and agent responses to corrections in real Arena agent sessions. See how leaderboard scores relate to median cost per task.
Limits on extrapolation: Performance and cost in other agent clients, direct API calls, fixed task sets, and individual task categories.
Model version: Gemini 4 Neon (High).
Test environment or client: Arena's Agent Mode, with sessions drawn from its global user community.
Reasoning tier and parameters: The leaderboard lists the tier as High. The exact model endpoint, system prompt, reasoning budget, tool configuration, and sampling settings are unspecified.
Arena used about 4,200 real agent sessions to calculate the index score for Neon (High). The cost chart plots index score against median cost per task. Its horizontal axis specifies that the cost figures are based on real user tasks from the previous 14 days. Arena's original post · Cost chart
According to the leaderboard FAQ, Agent Mode randomly assigns sessions to different models. The models decide when to use tools such as Bash and when to search the web or write files. Arena extracts behavioral signals from individual session logs, compares each model with the average baseline produced by random assignment, and combines the relative gains for each signal into a weighted score. New sessions and changes to the set of models in the evaluation can affect scores. Agent Arena leaderboard and FAQ
Neon's score of 1,336 was 33 points ahead of Carbon (High), 65 ahead of Hydrogen (High), and 107 ahead of Argon (High). The intervals for Neon and Claude Opus 5.5 (Max) overlap on the leaderboard chart. The chart does not specify how Arena calculated those intervals. Leaderboard chart
Arena placed Neon, Carbon, and Argon on the Pareto frontier for cost and score. Neon's median cost per task was 53% lower than that of Opus 5.5 (Max) and 69% lower than that of first-place Fable 5.2 (Max). Carbon's median cost in the snapshot was $1.34, about half of Neon's. Reply with cost figures · Cost chart
| Model and tier | Overall rank | Index score | Median cost per task (USD) |
|---|---|---|---|
| Claude Fable 5.2 (Max) | 1 | 1,453 | Exact figure unspecified |
| Gemini 4 Neon (High) | 2 | 1,336 | 2.82 |
| Claude Opus 5.5 (Max) | 3 | 1,335 | Exact figure unspecified |
| Gemini 4 Carbon (High) | 4 | 1,303 | 1.34 |
| Gemini 4 Hydrogen (High) | 9 | 1,271 | Exact figure unspecified |
| Gemini 4 Argon (High) | 11 | 1,229 | 0.69 |
Ranks and scores come from Arena's published leaderboard chart. Exact cost figures come from a reply in the same thread. Leaderboard chart · Reply with cost figures
| User signal for Neon (High) | Rank | Score |
|---|---|---|
| User praise and complaints about outcomes (Praise vs. Complaint) | 2 | 1,539 |
| Response to user corrections (Steerability) | 2 | 1,298 |
| User confirmation of successful task completion (Task Completion) | 3 | 1,316 |
All three category rankings and scores come from Arena's original post. Arena's original post
The ranks and score gaps reflect Arena's sample at publication. At the time of collection, the live leaderboard dated 2026-11-19 covered 68 models and 3,093,488 sessions. Neon (High) ranked third with 1,333 points and a median cost of $2.60 per task. Use the same snapshot when comparing rank and cost. Live Agent Arena leaderboard
The roughly 4,200 sessions are Neon's session count at publication. The original post leaves the total number of tasks, the mix of task types, and the sample sizes for each category ranking unspecified. Arena reports 1,316 as a Task Completion index score, which cannot be converted into a task success rate. The $2.82 median cost per task applies to Arena's published snapshot and cannot be used to estimate costs in other clients or API workflows. Arena's original post
Check the original X post for the ranks, sample size, and category results at publication, and the reply and charts in the same thread for the cost figures. The public materials omit the per-task prompts, tool call traces, and detailed user feedback needed to rerun the same index. Arena's original post · Reply with cost figures
Gemini 4 Neon