Latent Space publishes Frontier AEO tracker ranking agentic model preferences across 161 categories

Latent Space’s Astra project released a Frontier AEO tracker that ranks how frontier models choose tools and products for agentic workflows across 161 categories. The tracker was developed after "billions of tokens of prototyping, aligning, and scaling pipelines," and extends methodology from prior work on Claude Code’s choices.

The methodology runs six prompt variations over seven models, covering categories from coding agents to AI podcasts, AI sandboxes, managed databases, ASR models, and niche domains like angel investors and corporate spend. Answer extraction was performed by Astra and scored using a proprietary AEO metric that weights first choices, alternative choices, mentions, and penalizes mild or strong anti-recommendations.

The tracker identifies "soft biases" in coding agent recommendations, such as Fable and Opus (Claude Code) favoring their own ecosystems, Sol and Astra (Codex) favoring Codex and Grok favoring Cursor, Muse favoring Muse Code, and SWE-1.7 favoring Devin. Despite these biases, there are notable examples of GPT models recommending Claude, which the authors describe as "laudable nonbias."

Across the 161 categories, 28 have a universally dominant primary choice among all surveyed frontier models. The authors also highlight "close contests" and "always the vibe, never the vibe" categories as key battlegrounds for AEO. The dataset includes inspectable prompt-answer pairs, allowing users to examine the raw inputs and outputs for transparency.

One of the most surprising findings is the shift in source reliance and confidence between model generations. Anthropic’s Sol used a median of nine sources, while Astra used a median of five, and Opus used eleven sources, with Fable using fifteen. Astra was described as more "confident" or "efficient," being far less likely to change its answer when questions were lightly paraphrased. This suggests Astra may be converging on a smaller set of trusted sources and refusing to look further, a trend the authors note warrants deeper analysis across categories.

The authors also analyzed the sources influencing agent recommendations and identified failures that discourage models from reading content, such as markdown content-negotiation practices measured by Ora and Vercel. While the sample size for source analysis is small and limited to scraped tool call attempts, the authors validate that these AEO practices have real-world effects on model behavior.

The project acknowledges limitations, including the exclusion of Gemini/Antigravity, GLM/Zcode, and DeepSeek/DeepCode due to errors and rate limits. The authors invite feedback and collaboration to improve coverage, particularly from representatives of excluded model families.