GLM-52 897 —
GPT-56SC 873 —
CL-OP5X 865 -0.9%
GROK-46H 865 -0.9%
GEM-37FH 865 -0.9%
GPT-56T 861 —
GLM-5 856 —
MUSE-SPK 841 —
QWEN-38X 824 -2.3%
GPT-6A 820 —
KIMI-K3X 810 -1%
CL-FAB5H 787 -0.9%
CL-OP5H 764 -0.9%
CL-OP46H 742 -0.9%
CL-OP47H 733 -1.1%
GEM-38FH 676 -1%
CL-OP47 585 -0.7%
INKL 531 —
CL-OP46 496 -0.2%
CL-OP48 490 -0.2%
GLM-52 897 —
GPT-56SC 873 —
CL-OP5X 865 -0.9%
GROK-46H 865 -0.9%
GEM-37FH 865 -0.9%
GPT-56T 861 —
GLM-5 856 —
MUSE-SPK 841 —
QWEN-38X 824 -2.3%
GPT-6A 820 —
KIMI-K3X 810 -1%
CL-FAB5H 787 -0.9%
CL-OP5H 764 -0.9%
CL-OP46H 742 -0.9%
CL-OP47H 733 -1.1%
GEM-38FH 676 -1%
CL-OP47 585 -0.7%
INKL 531 —
CL-OP46 496 -0.2%
CL-OP48 490 -0.2%
← Back to feed

Artificial Analysis Benchmarks 7 Search APIs: Parallel Search Leads, Tavily Costs 9x More for Worse Results

Artificial Analysis released the AA Search Index on August 18, benchmarking 11 configurations across 7 search API providers on three tasks that measure how well search augments a base language model on real research problems. The verdict: search roughly doubles model accuracy, but the quality gap between providers is wide, and cost does not track quality.

The benchmark

The AA Search Index is the equal-weighted mean of three sub-benchmarks:

  • DeepSearchQA: Deep, multi-hop factual research queries
  • BrowseComp: Complex research that requires reading and synthesising web pages
  • AA-Omniscience: Proprietary AA knowledge-intensive benchmark

All providers were tested using the same underlying model (GPT-5.6 Luna, medium) in the same agent harness. The baseline — the same model without any search — scores 33 overall.

Results

ProviderOverallDeepSearchQABrowseCompAA-OmniSearch cost/task
Parallel Search (advanced)75817767$47.93
Exa Search (auto)74787470$65.57
Firecrawl Search73747473$30.48
Parallel Search (basic)73797368$45.14
Exa Search (fast)68766169$78.11
You.com Search68637466$68.93
Parallel Search (turbo)67707556$13.64
Keenable Search (pro)67706665$23.69
Keenable Search (realtime)67706466$26.87
Tavily Search (basic)66745964$126
Brave Search65626765$71.61
Model only (baseline)33451738—

What stands out

BrowseComp is where search matters most. The baseline model scores 17 on BrowseComp without any search access. The top providers push that to 75-77. This benchmark requires actually reading and synthesising retrieved pages — tasks where retrieval quality directly limits the model.

Tavily’s cost-quality ratio is severe. At $126 per task, Tavily is 9x the cost of Parallel Search (turbo) and still ranks 10th overall (66). It scores 74 on DeepSearchQA but falls apart on BrowseComp (59) and AA-Omniscience (64).

Firecrawl tops AA-Omniscience at 73, beating providers that rank above it overall. For knowledge-intensive research tasks specifically, Firecrawl’s web scraping approach outperforms the pure search-index approaches.

Budget option: Parallel Search (turbo) at $13.64. Scores 67 overall — 2 points below Exa (auto) at nearly 5x the cost. It leads all providers on BrowseComp (75). For cost-constrained agentic pipelines, turbo mode is a serious option.

Context

AA has been building out its agent-focused benchmark suite, alongside AA-Briefcase (knowledge work), Optima (custom benchmarks), and AgentPerf (hardware). The Search Index fills the gap for teams choosing search API providers for AI agent infrastructure.

Coverage launches with 7 providers; Artificial Analysis says it will expand.