vani(@vaniagrwall):I tested @OpenAI's GPT-6 Astra with @ExaAILabs, @KeenableAI, @p0, @ValyuOfficial and its built-in web search on 26 LiveBrowseComp questions. Exa matched Astra’s built-in web search on accuracy (11/26) at roughly 1/3 the estimated cost. It also finished faster. Keenable finished 1 behind (10/26). Native search made 513 tool calls versus Exa’s 650, but logged 56.5M input tokens versus 12.1M, including cached tokens. After accounting for cache pricing and search charges, estimated cost per question was $3.84 with native search versus $1.26 with Exa. results + methodology:

2026.09.08 04:55

I tested @OpenAI's GPT-6 Astra with @ExaAILabs, @KeenableAI, @p0, @ValyuOfficial and its built-in web search on 26 LiveBrowseComp questions. Exa matched Astra’s built-in web search on accuracy (11/26) at roughly 1/3 the estimated cost. It also finished faster. Keenable finished 1 behind (10/26). Native search made 513 tool calls versus Exa’s 650, but logged 56.5M input tokens versus 12.1M, including cached tokens. After accounting for cache pricing and search charges, estimated cost per question was $3.84 with native search versus $1.26 with Exa. results + methodology:

显示更多

转发到社区