Gemini 4 Neon scored 1700 Elo at Xhigh on AA-Briefcase v1.1. Its Presentation Elo of 1771 was the highest Artificial Analysis had measured when it published the post. Its overall Elo at the time still trailed Claude Sonnet 5.5 and Fable 5.2 at Max. Original post
What the benchmark can assess. It compares how models produce office briefs, spreadsheets, and slide decks with the same source materials and tools, with particular attention to how professionally the finished files are presented.
Limits on extrapolation. Accuracy on production tasks, adherence to a specific brand template, and native features in the Gemini web app or Google Slides.
Model version. Google Gemini 4 Neon. The post leaves the deployment snapshot unspecified.
Test environment or client. Artificial Analysis uses Stirrup to run agents in E2B sandboxes configured for each project week. Official methodology
Reasoning tiers and parameters. The post compares Neon at Xhigh, Carbon at High, and both Claude models at Max. The official leaderboard also lists Neon at Low, Medium, and High. The image leaves the reasoning tiers unlabeled, and the public materials leave the full request parameters for that example unspecified. Original post and image · Official leaderboard
Official scoring for AA-Briefcase v1.1 covers 91 tasks across 4 private projects. In each run, the model receives task instructions and access to source files, then submits its final file. The model completes each task independently, even when tasks in the same project share context. The output from one task is withheld from the next. Benchmark overview · Official methodology
Rubric scoring marks each criterion as pass or fail for a single output. Criteria include using the right source materials and reaching the correct conclusions. Analytical Quality and Presentation each compare pairs of outputs for the same task, judging the quality of the analysis and the professionalism of the presentation, respectively. The official leaderboard combines all three dimensions into AA-Briefcase Elo after converting rubric pass rates into comparable head-to-head results. Scoring methodology
Neon at Xhigh scored 89 Elo points higher overall than Carbon at High. At High, Neon scored 1632 and Carbon scored 1611, a gap of 21 points. The 89-point comparison in the post changes both the model and the reasoning tier. The post reports the largest gain in Presentation, up 170 Elo points to 1771. Original post · Official leaderboard · Presentation scores
| Model | Leaderboard reasoning tier | AA-Briefcase Elo |
|---|---|---|
| Claude Sonnet 5.5 | Max, Default Fallback | 1815 |
| Claude Fable 5.2 | Max, Default Fallback | 1785 |
| Gemini 4 Neon | Xhigh | 1700 |
| Gemini 4 Neon | High | 1632 |
| Gemini 4 Carbon | High | 1611 |
| Gemini 4 Neon | Medium | 1518 |
| Gemini 4 Neon | Low | 1297 |
The table uses the 2026-11-19 leaderboard snapshot and retains its reasoning-tier labels. Official leaderboard
The image in the original post shows two M&A due diligence slides from AA-Briefcase-Lite. The Neon slide uses a line chart, a column chart, and a cross-price elasticity matrix to explain the price premium for free-range eggs and its effect on profitability. The Claude Opus 5.5 slide uses a grouped column chart and a horizontal bar chart to explain consumer demand and price elasticity for different types of eggs. The slides let readers compare how the models organize information and choose charts. This public scenario falls outside the 91 tasks on the official leaderboard. Image from the original post · Public due diligence dataset
The score of 1771 reflects judges' pairwise comparisons of professional presentation across several file types. Before sharing a brief or slide deck with colleagues, check the figures, the sources, and whether the file meets the task requirements. The post shows one slide from each model. Assessing the full decks' quality and editability requires the original files. Benchmark overview · Image from the original post
To verify the reported scores, check the model names and reasoning tiers on the AA-Briefcase v1.1 leaderboard, then switch to Analytical Quality & Presentation Elo under Results. To rerun the public example, use the instructions, source files, and validation rules for task 3 in AA-Briefcase-Lite and submit target_assessment.pptx. The benchmark publisher released Stirrup and the Lite scenario. The four projects behind the official leaderboard remain private. Official methodology · Public task instructions · Stirrup
Gemini 4 Neon