Free Intensive Nov 21–22Free 2-day hands-on Vibecoding Intensive on Nov 21–22 at 16:00 MSK (Sat–Sun) · Live Cohort 3 starts Nov 23Join Free →Cohort 3 (Nov 23)
ReleaseAPI / Product · ActionableX: @ArtificialAnlys→ original

Parallel Search Tops 26-API AI Agent Search Ranking With 83 Points

If an AI agent searches the web for answers, the quality of its results depends on the search API it uses to reach the internet. Artificial Analysis ran 1,700 tasks through 26 such services. Parallel Search scored 83 on the composite index, tying Keenable Search (pro). The same model without search scores 33, so the service adds roughly 50 points. The fast mode delivers 82 points for $0.014 per task, against a ranking median of $0.063, which makes it the first option to test.

AI-processed from X: @ArtificialAnlys; edited by Hamidun News
Parallel Search Tops 26-API AI Agent Search Ranking With 83 Points
Source: X: @ArtificialAnlys. Collage: Hamidun News.
Listen to article
Article Fact Sheet & Key Figures
ParameterDetails
Primary sourceX: @ArtificialAnlys ↗
Editor-in-Chief
DeepSearchQAboth modes scored 89 F1, while Keenable Search (pro) scored 88
Pricing$1 per 1,000 requests in fast mode and $5 in advanced mode
Key metric #3Since August, advanced rose from 75 to 83 and fast rose from 73 to 82

Picture an agent tasked with finding out what a competitor's current terms are, or what changed in the latest version of a library. It doesn't know the answer on its own, so it sends a query to a search API and reads whatever comes back. If the API returns junk or misses the right page, even a strong model will answer inaccurately. That's why Artificial Analysis's fresh benchmark matters to anyone choosing such a service for production: 26 search APIs were tested under identical conditions, on 1,700 tasks that require finding information on the web.

Parallel Search took the lead. Its advanced mode scored 83 on the composite Artificial Analysis Search Index and ranked first out of 26. The fast mode scored 82 and placed fourth. For scale: the same model without any search scores 33, so good web search adds about 50 points to an agent's performance.

How the Benchmark Was Run

The analysts used their own Stirrup agent harness and connected Parallel Search as a web_search tool. The model was GPT-5.6 Luna with medium reasoning effort. On each call, the model fills in the fields itself: an optional objective, where it describes the goal of the search in its own words, and search_queries, meaning one to three search queries. The agent decides how to phrase the search, and the service assembles the results.

The index standings look like this. Advanced sits level with Keenable Search (pro), which also scores 83. It leads Exa (auto, dynamic highlights) and Perplexity (medium) by three points, and both of those score 80. The fast mode, at 82, is also slightly ahead of that pair.

Individual Tests and Trends

Parallel looks strongest on DeepSearchQA. Both modes scored 89 on the F1 metric, taking first and second place. Keenable Search (pro) scored 88 here. F1 measures the balance between completeness and precision: the service finds what's needed without mixing in irrelevant material.

Advanced also ranks first on AA-Omniscience with 74%, while Keenable Search (pro) scored 73%. On BrowseComp it ranks third with 87%, level with Keenable Search (pro) and Perplexity (medium).

The service's own growth is also notable. In the August runs, advanced scored 75 and fast scored 73. In the November runs, those figures are 83 and 82, an increase of 8 to 9 points. Search APIs change quickly, so comparisons from six months ago are already a poor basis for choosing.

The Money

Fast mode costs about $0.014 per task. Only TinyFish, at $0.009, is cheaper, and the ranking median is $0.063, which makes fast roughly 4.5 times cheaper than the typical entrant. Advanced costs about $0.030 per task and ranks sixth for affordability.

Parallel's pricing is $1 per 1,000 requests in fast mode and $5 per 1,000 requests in advanced mode. The per-task cost in the benchmark and the per-request price differ: within a single task, an agent may call search several times.

What to Do in Practice

If your agent currently reaches the web through Exa or Perplexity, it makes sense to run an A/B test with Parallel. Starting with fast is the logical move: it trails advanced by only one point and costs noticeably less. Advanced is worth enabling where factual accuracy and multi-step questions matter most.

The caveats: all the numbers come from one model and one harness, and the benchmark was run by Artificial Analysis itself. In a different framework or with a different LLM, the standings may shift somewhat. The gap between the leaders, 2 to 3 points, is small, and Parallel's first place is shared with Keenable Search (pro). So the ranking is a starting point, and the final choice is best made on your own queries: take typical user questions, run them through two or three APIs, and compare answer quality and total cost.

Story

Key Questions Answered

How the Benchmark Was Run?

The analysts used their own Stirrup agent harness and connected Parallel Search as a web_search tool. The model was GPT-5.6 Luna with medium reasoning effort. On each call, the model fills in the fields itself: an optional objective, where it describes the goal of the search in its own words, and search_queries, meaning one to three search queries. The agent decides how to phrase the search, and the service assembles the results. The index standings look like this.

What are the key details on Individual Tests and Trends?

Parallel looks strongest on DeepSearchQA. Both modes scored 89 on the F1 metric, taking first and second place. Keenable Search (pro) scored 88 here. F1 measures the balance between completeness and precision: the service finds what's needed without mixing in irrelevant material. Advanced also ranks first on AA-Omniscience with 74%, while Keenable Search (pro) scored 73%. On BrowseComp it ranks third with 87%, level with Keenable Search (pro) and Perplexity (medium). The service's own growth is also notable.

What are the key details on The Money?

Fast mode costs about $0.014 per task. Only TinyFish, at $0.009, is cheaper, and the ranking median is $0.063, which makes fast roughly 4.5 times cheaper than the typical entrant. Advanced costs about $0.030 per task and ranks sixth for affordability. Parallel's pricing is $1 per 1,000 requests in fast mode and $5 per 1,000 requests in advanced mode. The per-task cost in the benchmark and the per-request price differ: within a single task, an agent may call search several times.

ZK
Zhemal Khamidun — CPO at Alpina Digital, product manager of AlpinaGPT. Building AI products since 2019. Hamidun News is a pet project that grew into a production aggregator with 6,000+ readers.
✦ Private AI Builders Community · Vibe Club

Ship AI Faster with Founders, Engineers & Weekly Architecture Reviews

Inside Vibe Club (hamidun.com/club) — weekly live masterminds with Zhemal Khamidun, Tech Radar, full LMS access, and 370+ production Claude Code skills.

Join Vibe Club →Explore Hamidun Academy
What do you think?
Loading comments…