Keenable Search 1.0 Tops AI Search Rankings at 37% Below Median Cost

Artificial Analysis has updated its Search API leaderboard, placing Keenable Search 1.0 first and second. Pro mode scored 83 on the Artificial Analysis Search Index, while realtime mode scored 82. Exa with auto search and dynamic highlights, and Perplexity at medium effort, each scored 80.

The results combine higher benchmark scores with below-median cost and latency. That mix matters for developers building agents that may issue several searches before producing an answer, because small differences in request price and response time compound across each task.

The lead spans three tests

The index averages three public benchmarks with equal weight. Each measures a different part of web-grounded question answering:

Both Keenable modes scored 88 on DeepSearchQA, taking the top two positions among 26 tested configurations. Exa in auto mode scored 86. On BrowseComp, Keenable pro tied Perplexity medium at 87%. Its 73% AA-Omniscience result matched Exa with auto search and dynamic highlights, as well as Firecrawl.

The fixed answer model scored 33 without search, compared with Keenable pro’s index score of 83. The resulting 50-point gain is the largest search-related uplift currently shown on the leaderboard. Artificial Analysis provides further scoring details in its benchmark methodology .

Lower spend, shorter waits

Keenable charges $4 per 1,000 search requests in both modes. Artificial Analysis estimates a total cost of about $0.040 per benchmark task, including roughly $0.008 for model inference. Those figures imply about $0.032 in search spend, or approximately eight requests per task at the listed rate.

The leaderboard’s median task cost is about $0.063. Keenable therefore costs roughly 37% less than the median in this test, with search requests accounting for about 80% of its total task cost.

Realtime ranked second for task latency, followed by pro in third. Octen led at 14.6 seconds. Individual Keenable calls returned in about 330 milliseconds for realtime and 390 milliseconds for pro.

The leaderboard separately reports a median of 12.5 seconds spent inside web_search across an entire task. That figure represents cumulative search time, rather than the latency of one API call.

Both Keenable modes sit on the benchmark’s quality-to-cost and quality-to-latency Pareto frontiers. Under this definition, no tested configuration produced a higher score at the same or lower cost, and none produced a higher score in less time.

A controlled agent loop

Artificial Analysis held the answer model and agent environment constant so that each run primarily measured the search provider. The tests used GPT-5.6 Luna at medium reasoning effort, a 25-turn limit, and up to 10 results per search.

Keenable’s search cards were disabled with disable_cards=true . That setting isolates retrieval and ranking from answer text or other formatting prepared by the provider.

The agent ran in Stirrup , the organization’s open-source evaluation harness. It could call two research tools:

• web_search , backed by the provider under evaluation
• web_fetch , used to retrieve individual URLs

Once the model had enough evidence, it called a finish tool and submitted its answer. A run that exhausted all 25 turns without calling finish received a score of zero. This design captures retrieval quality alongside the search behavior it induces, including repeated queries and failed completion attempts.

The leaderboard identifies Keenable’s configurations as pro and realtime, but the published setup does not explain how their retrieval strategies differ. Developers comparing the modes therefore have performance measurements without corresponding implementation details.

What developers can take from the results

Provider selection affects more than result freshness or crawl coverage. Ranking, snippet structure, result formatting, and relevance can change how many searches an agent performs and whether it reaches a supported answer within its turn budget.

• For multi-step research: Keenable pro and Perplexity medium tied at 87% on BrowseComp.
• For interactive agents: Keenable’s reported per-call latency stayed below 400 milliseconds, while end-to-end tasks averaged about 15 seconds.
• For high-volume workloads: Keenable’s estimated $0.040 task cost was below the leaderboard median of $0.063.
• For evaluation: teams should test providers with their own prompts, domains, turn limits, and answer model before choosing a production default.

These rankings describe one benchmark suite, one agent harness, and one fixed answer model. They provide a controlled comparison of search APIs under that setup, while production results will also depend on query patterns, geographic coverage, caching, rate limits, failure handling, and the model orchestrating the searches.