Claude Fable 5 + Exa scored 48.5% overall, ahead of the same model using native search at 46.9%.
The only variable Vals changed was the search layer.
The only variable Vals changed was the search layer.
Jul 23, 2026 · 6:13 PM UTC
1
0
5
106
The difference was clearest in finance. After controlling for task difficulty, agents using Exa scored 6.51 points higher than agents using native search.
1
0
4
55
The biggest gains came from tasks where finding the right underlying information matters most: Market Analysis: 54% vs. 36% Earnings Analysis: 73% vs. 59% Adjustments: 44% vs. 36%
1
0
4
52
This is why search is not just another tool call. As models get better, the quality of the context they can retrieve increasingly determines the quality of the work they produce.
Read the full Web Search Index from Vals AI: vals.ai/benchmarks/web…
Read the full Web Search Index from Vals AI: vals.ai/benchmarks/web…
0
0
4
35