Three web search APIs lead on accuracy for AI agents, Parallel, Exa and cloro, and the order changes with the question type. In an open benchmark of six web search APIs on 2,182 questions, each API got one search per question, and the same model answered from the results.
- Short factual questions: Parallel (97.8%), Exa (97.3%) and cloro (96.9%) are within the margin of error of each other. SerpApi, Brave and Tavily score 91.8% to 94.7%.
- Questions about recent events: the two Google SERP APIs lead, SerpApi at 88.5% and cloro at 88.3%. Exa and Parallel score 83.0% and 82.7%.
- Multi-hop questions: cloro leads at 80.2%, ahead of SerpApi at 77.4%. Parallel (65.3%) and Exa (64.0%) are 15 points behind, Brave (53.9%) and Tavily (48.4%) further.
Which web search API is the most accurate for AI agents?
Accuracy is the share of questions answered correctly, with the 95% interval in brackets. Each column is one question type.
| Search API | Short factual: SimpleQA Verified (1,000) | Recent events: FreshQA (358) | Multi-hop: FRAMES (824) |
|---|---|---|---|
| Parallel | 97.8% (96.7-98.5) | 82.7% (78.4-86.2) | 65.3% (62.0-68.5) |
| Exa | 97.3% (96.1-98.1) | 83.0% (78.7-86.5) | 64.0% (60.6-67.2) |
| cloro | 96.9% (95.6-97.8) | 88.3% (84.5-91.2) | 80.2% (77.4-82.8) |
| SerpApi | 94.7% (93.1-95.9) | 88.5% (84.8-91.4) | 77.4% (74.4-80.2) |
| Brave Search API | 93.1% (91.4-94.5) | 84.9% (80.8-88.3) | 53.9% (50.5-57.3) |
| Tavily | 91.8% (89.9-93.3) | 81.8% (77.5-85.5) | 48.4% (45.0-51.8) |
No API is first in every column. So the right question for an agent is which kind of question it asks most.
Which API is most accurate on short factual questions?
SimpleQA Verified has 1,000 short questions with one checkable answer. Parallel scored 97.8%, Exa 97.3% and the top Google SERP API 96.9%. A paired test on the same questions finds no significant gap between the three (p = 0.15 against Parallel, p = 0.63 against Exa). All three were more accurate than SerpApi (94.7%), Brave (93.1%) and Tavily (91.8%).
If your agent mostly looks up single facts, accuracy does not decide between Parallel, Exa and the Google SERP API at the top. Cost and latency decide.
Which API is most accurate on questions about recent events?
FreshQA has 358 questions whose answers change over time. A third of them carry a false premise that the model must point out. SerpApi (88.5%) and cloro (88.3%) tied. Both Google SERP APIs were more accurate than Tavily (81.8%), Exa (83.0%) and Parallel (82.7%), each at p of 0.0012 or lower. The gap to Brave (84.9%) was not significant (p = 0.065).
For news, prices, rankings or facts that changed last week, the two Google SERP APIs answered more questions than Tavily, Exa and Parallel.
Which API is most accurate on multi-hop questions?
FRAMES has 824 questions that each need facts from two or more Wikipedia pages. The top Google SERP API answered 80.2% correctly, ahead of SerpApi at 77.4% in a paired test on the same questions (p = 0.022). Parallel scored 65.3%, Exa 64.0%, Brave 53.9% and Tavily 48.4%.
The gap comes from coverage. Parallel, Exa, Brave and Tavily replied "not found" on 151 to 295 of the 824 questions, because their 10 results did not hold every fact the question needs. When they did answer, five of the six APIs were right 78% to 81% of the time. For an agent, each "not found" is a second search, a page fetch or a failed task.
Exa and Parallel are statistically tied on all three datasets. Exa beat Tavily on factual and multi-hop questions and tied with it on recent events.
Which part of a search result makes the answer correct?
A second test kept the saved Google responses for 100 SimpleQA Verified questions and changed only what the model could read.
| What the model read | Accuracy | Tokens per question |
|---|---|---|
| AI Overview, then 10 results with title, URL and snippet | 99% | 843 |
| AI Overview and snippets only | 99% | 528 |
| AI Overview and the top 5 results | 99% | 549 |
| AI Overview alone | 95% | 174 |
| 10 results without the AI Overview | 86% | 747 |
| 10 results without the AI Overview, plus the text of the top result pages | 91% | 6,239 |
The AI Overview carries most of the accuracy. Without it, the model lost 13 questions and gained none (p below 0.001). The snippets carry most of the rest. Titles, URLs, a second results page and the full JSON response added tokens and no accuracy. Without the AI Overview, page text raised accuracy from 86% to 91%, which is not significant, and it read 7 times as many tokens as the 99% format.
For a RAG pipeline, this gives a short context format: the AI Overview first, then one snippet per result. Keep the URLs only if the agent must cite its sources. These tests used 100 questions, so they confirm large effects only.
What do speed and cost add to the choice?
Exa was the fastest at 0.74 seconds per search at the median. Brave took 1.21 seconds, Tavily 2.00 and Parallel 2.99. The two Google SERP APIs were the slowest: 7.96 seconds and 8.20 seconds for SerpApi. If your agent must answer in under 3 seconds, Exa, Brave, Tavily or Parallel is the right choice today.
On cost per correct answer, the order is different. On SimpleQA Verified, cloro cost $1.98 per 1,000 correct answers, the lowest of the six, against $5.11 for Parallel, $5.37 for Brave, $5.45 for Tavily, $7.19 for Exa and $15.08 for SerpApi, which needed about two searches per question. List prices were read on November 5, 2026.
How was accuracy measured?
Every API got the same setup:
- One search per question. The query is the question text, with no rewrite, no second search and no page fetch.
-
10 results in the API's default search mode. Answer features sold as separate options (Tavily
include_answer, Exasummary) were off. -
The same reader. One model (
gpt-5.6-luna) read the results and answered from them alone, or replied "NOT FOUND". -
The same grader. One model (
gpt-5.6-sol) graded each answer with the unchanged OpenAI simple-evals grader prompt. A second grader agreed on 99.7% of the 6,100 SimpleQA answers it checked. - The same questions, shuffled with seed 42. The runs were on November 4, 5, 10 and 11, 2026.
The datasets are public: SimpleQA Verified, FreshQA and FRAMES.
What does this benchmark not show?
- One search measures the search API, not the agent. An agent that searches many times and opens pages can get a high score from a weaker search API. FRAMES is usually run that way.
- Some pages copy benchmark answers. On SimpleQA, 193 of 1,000 questions had a result that names the dataset. Without them, the order of the APIs does not change.
- Answer keys go stale. On 30 FreshQA questions, most APIs disagree with an answer key that may be out of date. Without them, the order does not change.
This benchmark is published by cloro, which sells one of the six APIs tested. For that reason, the harness, the settings and all 13,092 graded answers are public in the benchmark repo. If a setting is unfair to an API, change it and run it again. The head-to-head tests, the cost per search and the full method are in the web search API benchmark write-up.
How do you test the APIs on your own questions?
The order changed between the three datasets, so the best test is your own questions. The harness is one Python file with no packages to install. Set OPENAI_API_KEY and the keys of the APIs you want to compare, then run 50 to 100 of your questions:
python3 benchmark.py --providers exa,parallel,tavily,brave --n 100 --workers 4
Run --n 3 first to find a broken key before you spend the budget.
Top comments (0)