Keenable Beats Every Search Tool on Vals' Finance and Legal AI Index

Keenable in Pro mode, paired with GPT-5.6 Sol, now leads the Vals index with a score of 49.08%. The benchmark measures whether an AI agent can complete expert-written finance and legal tasks using a specific web search integration.

Vals gives each agent the evaluated tool as its sole browsing interface. Within a model comparison, the underlying model and agent harness remain fixed while the search integration changes. The overall leaderboard still mixes models, so each rank reflects a complete model-tool configuration rather than search quality alone.

Less than one point separates the leaders

GPT-5.6 Sol with Keenable ranks first overall, followed by Claude Fable 5 with Exa at 48.45% and Claude Fable 5 with native search at 46.94%. Keenable scores 53.9% on finance and 44.2% on legal, where Claude Fable 5 with native search remains the strongest configuration.

The 0.63 percentage-point gap between the top two configurations is narrow. The reported scores alone do not establish whether that difference is statistically significant, so sample size and uncertainty should inform any claim about a definitive lead.

Model tokens set the price

GPT-5.6 Sol with Keenable costs an average of $8.84 per task, including $0.41 for search. Model tokens account for the remaining $8.43, or about 95% of the total. Search API pricing therefore has less effect on the bill than model selection, context volume, and agent behavior.

GPT-5.6 Sol with Exa illustrates the effect of uncontrolled tool use. Its mean cost reaches $18.97 because several tasks triggered hundreds of search calls before the model produced an answer. Those runs raise the average and show how agent loops can overwhelm nominal per-query prices.

Keenable puts cited sources near the top

Vals’ source-position data offers one explanation for Keenable’s score. Sources that the agent eventually cited appeared as Keenable’s first result 80% of the time in finance and 69% in legal. They appeared within the top three results 92% and 85% of the time, respectively.

The same runs show GPT-5.6 Sol searching more often with Keenable than with Exa. That difference makes the source ranking and call volume difficult to separate: stronger retrieval may improve accuracy, and additional searches may also contribute.

SEC filings account for 77% of Keenable’s finance citations, compared with roughly 54% to 58% for native search and Exa. Keenable posts its largest category gains in Market Analysis, Adjustments, and Earnings Analysis for finance, plus Resolving Conflicting Precedent and Cross-Jurisdiction Comparison for legal.

Legal results and cost spikes limit the win

Claude Fable 5 with native search still leads the legal domain. GPT-5.6 Sol with native search also presents a lower-cost option overall, scoring 43.58% at $0.95 per task. Keenable improves that model’s score by 5.48 percentage points while increasing average cost by $7.89.

Average costs conceal a long tail. A handful of Keenable runs reached about $190 in finance and $281 in legal, creating substantial exposure for systems that process tasks without spending controls.

Production implementations can contain that exposure with per-task call limits, token budgets, timeouts, loop detection, and explicit fallback behavior. Logging query counts and cumulative model tokens also makes abnormal runs easier to stop before they exhaust a budget.

Native search saves tokens by compressing results

OpenAI’s native search processes results server-side and gives the model a compact result block. External integrations can place more retrieved content into the model’s context, increasing input-token charges. That mechanical difference helps explain the $0.95 average cost for GPT-5.6 Sol with native search.

The benchmark establishes an association between compact native results, lower cost, and lower reported accuracy for this configuration. It does not isolate result compression as the cause because ranking quality, retrieved content, and agent search behavior also differ.

A benchmark built around completed work

Vals designed the index to reduce the contamination risk affecting static, public search benchmarks. Its held-out test set uses expert-written professional tasks and evaluates the final answer produced with each search integration.

BrowseComp illustrates the concern. Vals cites a 2026 study in which a frontier model answered 44.5% of BrowseComp questions without search tools, indicating that model knowledge can account for a substantial share of performance. Holding the model and harness fixed gives the Vals method a clearer view of how each search integration changes task results.

Choosing a configuration

• Highest reported score: GPT-5.6 Sol with Keenable Pro reaches 49.08% at an average cost of $8.84 per task.
• Lower-cost near leader: Claude Fable 5 with Exa reaches 48.45% at $3.48 per task.
• Lowest listed cost: GPT-5.6 Sol with native search reaches 43.6% at $0.95 per task.

External search APIs produce higher reported scores than native search in the matched configurations shown here, while model usage remains the largest cost component. Production selection should also account for task-specific accuracy, latency, source freshness, API reliability, and worst-case spending, none of which can be resolved by the aggregate leaderboard alone.