Keenable Tops Vals Web Search Index, With a Steep Cost Tradeoff

Vals AI built its Web Search Index to ask whether an agent equipped with a given search tool can answer expert-written professional questions. It currently covers finance, with questions written by investment banking and private equity analysts, and legal research, with questions and grading standards set by licensed attorneys; on the leaderboard, LLM judges score answers against those expert rubrics. To isolate what the search tool contributes, Vals holds the model and harness fixed and swaps only the search tool, then grades the agent's final answer rather than the search results. Vals says agents with search score 30 to 50% on both domains, while a run with no search scored 7.4% on finance and 2.9% on legal.

Vals reports that Keenable, which its post describes as running in Pro mode, paired with OpenAI's GPT-5.6 Sol now ranks first overall at 49.1%, leading finance at 53.9% and placing second in legal at 44.2%, behind Anthropic's Claude Fable 5 with its native search. The leaderboard, updated October 21, lists the run simply as Keenable and shows how narrow the lead is and what it costs. Claude Fable 5 with Exa is second at 48.45% for $3.48 per task, and Fable 5 with native search is third at 46.94% for $4.37. The Keenable run averages $8.84 per task, about two and a half times the Exa pairing, and Keenable's own search fees are $0.41 of that. For comparison, the same GPT-5.6 Sol model with its native search scores 43.58% for $0.95.

Vals notes two more limits. The cost per task is uneven: the median Keenable task costs $3.22, but the most expensive tasks reach about $190 in finance and $281 in legal. And Keenable has so far been run only with GPT-5.6 Sol, so the index cannot yet say whether it would also top the table with another model. On the index's own terms, Keenable adds about five and a half points to GPT-5.6 Sol over native search, at roughly nine times the average task cost.