Jul 10, 2026
- [How Nooks cut web search costs 70.5% by switching to Parallel](https://parallel.ai/blog/case-study-nooks)
Tags:Customers
Author: By ParallelJuly 13, 2026
# Introducing Parallel Search Turbo
Reading time: 3 min
Today, we're releasing **Turbo **mode for Parallel Search[Parallel Search](/products/search), the fastest and most affordable way to ground your agents in high-quality information from the web.
## Do 14x more research in 5x less time
With a median latency of 200ms (p50) and a price of just $1 per 1,000 requests, Turbo is in a class of its own. It’s up to 14x cheaper than the default search in frontier models, while maintaining similar or better accuracy.
| Parallel Search Turbo | Fast Search APIs | Frontier Model Search | SERP APIs | |
|---|---|---|---|---|
| Cost / 1K Requests | $1 | $5-7 | $10-14 | $1 |
| Median Latency | 200ms | 200-350ms | ~1s | ~1s |
| Output | LLM-Ready Results | LLM-Ready Results | LLM-Ready Results | Raw Search Results |
## New use cases unlocked by low-latency, low-cost search
Turbo is built for the workloads where latency is core to the product experience: voice agents, chat, support, anywhere a user is waiting for an answer.
Voice and chat apps have previously had to make compromises on search quality. Parallel Search Turbo eliminates the tradeoff with the added benefit of being highly cost-effective.
- - **Voice AI:** Voice agents need to be low-latency to feel magical. Use Turbo with real-time voice models to add web grounding without introducing awkward pauses.
- - **Deep Research: **Applications that use deep research for decision-making can now perform significantly more searches, broadening the surface in comprehensive fan-outs.
- - **Consumer chat**: Consumer chat applications rely on heuristics to decide when web grounding can enhance a response. Apps can now use Turbo’s low latency and cost to offer grounded search as the default user experience.
- - **Reinforcement learning**: Training models with web grounding depends on millions of live queries per rollout. Turbo cuts that latency so more of your compute budget goes to learning.
- - **Small, local models: **Small models can answer harder questions with generous use of web search, making efficient local AI more viable.
BrowseCompHLEWebWalkerSimpleQACoding
Accuracy (%)
Parallel Turbo51% / 216ms
Brave Search38.3% / 430ms
Exa Instant33.7% / 361ms
SerpAPI23.3% / 999ms
Tavily Ultra Fast19.3% / 357ms
P50 SEARCH LATENCY (ms)
ACCURACY (%)
Loading chart...
Latency: p50 client-side wall clock per search API request, in ms, shown on a log scale (best across runs). OpenAI Web Search is omitted (single search-call latency not available).
Parallel
Others
Benchmark comparison across p50 Search Latency (ms) and Accuracy (%). Latency: p50 client-side wall clock per search API request, in ms, shown on a log scale (best across runs). OpenAI Web Search is omitted (single search-call latency not available).
+Methodology
**Dataset**
BrowseComp[BrowseComp](https://openai.com/index/browsecomp/), created by OpenAI, contains 1,266 questions that require persistent browsing to locate hard-to-find, entangled information on the web.
**Evaluation methodology**
Multi-hop evaluation: a GPT-5.4 agent runs with up to 20 tool calls (search_web, plus web_fetch for engines with an extract API: Parallel, Exa, and Tavily; Brave and SerpAPI are search-only). Answers are graded by an LLM judge (GPT-5.4, per-suite grader prompts). Benchmarks were run across multiple sessions, with the best observed scores selected for each provider.
Latency: search-call latency is the client-side wall clock measured around a single provider search API request, from a client in us-central; we report the p50 across all questions (best across runs). OpenAI Web Search scored 57.7% accuracy on this suite but is not plotted because its single search-call latency is not available.
**Testing dates**
Evals were run between July 10 and 12, 2026.
**Dataset**
BrowseComp[BrowseComp](https://openai.com/index/browsecomp/), created by OpenAI, contains 1,266 questions that require persistent browsing to locate hard-to-find, entangled information on the web.
**Evaluation methodology**
Multi-hop evaluation: a GPT-5.4 agent runs with up to 20 tool calls (search_web, plus web_fetch for engines with an extract API: Parallel, Exa, and Tavily; Brave and SerpAPI are search-only). Answers are graded by an LLM judge (GPT-5.4, per-suite grader prompts). Benchmarks were run across multiple sessions, with the best observed scores selected for each provider.
Latency: search-call latency is the client-side wall clock measured around a single provider search API request, from a client in us-central; we report the p50 across all questions (best across runs). OpenAI Web Search scored 57.7% accuracy on this suite but is not plotted because its single search-call latency is not available.
**Testing dates**
Evals were run between July 10 and 12, 2026.
### BrowseComp (Latency)
| Series | Model | p50 Search Latency (ms) | Accuracy (%) | | -------- | ----------------- | ----------------------- | ------------ | | Parallel | Parallel Turbo | 216 | 51 | | Others | Exa Instant | 361 | 33.7 | | Others | Tavily Ultra Fast | 357 | 19.3 | | Others | Brave Search | 430 | 38.3 | | Others | SerpAPI | 999 | 23.3 |
Latency: p50 client-side wall clock per search API request, in ms, shown on a log scale (best across runs). OpenAI Web Search is omitted (single search-call latency not available).
**Dataset**
BrowseComp[BrowseComp](https://openai.com/index/browsecomp/), created by OpenAI, contains 1,266 questions that require persistent browsing to locate hard-to-find, entangled information on the web.
**Evaluation methodology**
Multi-hop evaluation: a GPT-5.4 agent runs with up to 20 tool calls (search_web, plus web_fetch for engines with an extract API: Parallel, Exa, and Tavily; Brave and SerpAPI are search-only). Answers are graded by an LLM judge (GPT-5.4, per-suite grader prompts). Benchmarks were run across multiple sessions, with the best observed scores selected for each provider.
Latency: search-call latency is the client-side wall clock measured around a single provider search API request, from a client in us-central; we report the p50 across all questions (best across runs). OpenAI Web Search scored 57.7% accuracy on this suite but is not plotted because its single search-call latency is not available.
**Testing dates**
Evals were run between July 10 and 12, 2026.
### HLE (Latency)
| Series | Model | p50 Search Latency (ms) | Accuracy (%) | | -------- | ----------------- | ----------------------- | ------------ | | Parallel | Parallel Turbo | 220 | 52.7 | | Others | Exa Instant | 358 | 49.3 | | Others | Tavily Ultra Fast | 243 | 42 | | Others | Brave Search | 563 | 47.7 | | Others | SerpAPI | 865 | 40 |
Latency: p50 client-side wall clock per search API request, in ms, shown on a log scale (best across runs). OpenAI Web Search is omitted (single search-call latency not available).
**Dataset**
Humanity's Last Exam (HLE)[Humanity's Last Exam (HLE)](https://lastexam.ai/), created by CAIS and Scale AI, is a benchmark of expert-written questions at the frontier of human knowledge across dozens of subjects.
**Evaluation methodology**
Multi-hop evaluation: a GPT-5.4 agent runs with up to 20 tool calls (search_web, plus web_fetch for engines with an extract API: Parallel, Exa, and Tavily; Brave and SerpAPI are search-only). Answers are graded by an LLM judge (GPT-5.4, per-suite grader prompts).
Latency: search-call latency is the client-side wall clock measured around a single provider search API request, from a client in us-central; we report the p50 across all questions (best across runs). OpenAI Web Search scored 66% accuracy on this suite but is not plotted because its single search-call latency is not available.
**Testing dates**
Evals were run between July 10 and 12, 2026.
### WebWalker (Latency)
| Series | Model | p50 Search Latency (ms) | Accuracy (%) | | -------- | ----------------- | ----------------------- | ------------ | | Parallel | Parallel Turbo | 217 | 75.7 | | Others | Exa Instant | 336 | 65 | | Others | Tavily Ultra Fast | 240 | 63.7 | | Others | Brave Search | 503 | 65.7 | | Others | SerpAPI | 761 | 50.7 |
Latency: p50 client-side wall clock per search API request, in ms, shown on a log scale (best across runs). OpenAI Web Search is omitted (single search-call latency not available).
**Dataset**
WebWalkerQA[WebWalkerQA](https://huggingface.co/datasets/callanwu/WebWalkerQA) evaluates an agent's ability to traverse the web — navigating through linked pages to find information that a single search does not surface.
**Evaluation methodology**
Multi-hop evaluation: a GPT-5.4 agent runs with up to 20 tool calls (search_web, plus web_fetch for engines with an extract API: Parallel, Exa, and Tavily; Brave and SerpAPI are search-only). Answers are graded by an LLM judge (GPT-5.4, per-suite grader prompts).
Latency: search-call latency is the client-side wall clock measured around a single provider search API request, from a client in us-central; we report the p50 across all questions (best across runs). OpenAI Web Search scored 80.7% accuracy on this suite but is not plotted because its single search-call latency is not available.
**Testing dates**
Evals were run between July 10 and 12, 2026.
### SimpleQA (Latency)
| Series | Model | p50 Search Latency (ms) | Accuracy (%) | | -------- | ----------------- | ----------------------- | ------------ | | Parallel | Parallel Turbo | 240 | 91 | | Others | Exa Instant | 335 | 89.3 | | Others | Tavily Ultra Fast | 150 | 72 | | Others | Brave Search | 475 | 87 | | Others | SerpAPI | 652 | 76.7 |
Latency: p50 client-side wall clock per search API request, in ms, shown on a log scale (best across runs).
**Dataset**
SimpleQA[SimpleQA](https://openai.com/index/introducing-simpleqa/), created by OpenAI, contains 4,326 short, fact-seeking questions across a variety of domains.
**Evaluation methodology**
Single-step evaluation: the raw question is sent as the search query (num_results=10, with an equal ~1,000 character-per-result content budget for every engine) and GPT-5.4 (reasoning: high) synthesizes an answer from the search results only. Answers are graded by an LLM judge (GPT-5.4, per-suite grader prompts).
Latency: search-call latency is the client-side wall clock measured around a single provider search API request, from a client in us-central; we report the p50 across all questions (best across runs).
**Testing dates**
Evals were run between July 10 and 12, 2026.
### Coding (Latency)
| Series | Model | p50 Search Latency (ms) | Accuracy (%) | | -------- | ----------------- | ----------------------- | ------------ | | Parallel | Parallel Turbo | 216 | 79.7 | | Others | Exa Instant | 341 | 76.7 | | Others | Tavily Ultra Fast | 208 | 71.9 | | Others | Brave Search | 514 | 64.3 | | Others | SerpAPI | 683 | 54 |
Latency: p50 client-side wall clock per search API request, in ms, shown on a log scale (best across runs). OpenAI Web Search is omitted (single search-call latency not available).
**Dataset**
A proprietary coding dataset derived from production queries to Parallel's search API.
**Evaluation methodology**
Multi-hop evaluation: a GPT-5.4 agent runs with up to 20 tool calls (search_web, plus web_fetch for engines with an extract API: Parallel, Exa, and Tavily; Brave and SerpAPI are search-only). Answers are graded by an LLM judge (GPT-5.4, per-suite grader prompts).
Latency: search-call latency is the client-side wall clock measured around a single provider search API request, from a client in us-central; we report the p50 across all questions (best across runs). OpenAI Web Search scored 76.7% accuracy on this suite but is not plotted because its single search-call latency is not available.
**Testing dates**
Evals were run between July 10 and 12, 2026.
## Replace SERP APIs
Turbo is especially well-suited for replacing workflows built on SERP APIs. While SERP calls are affordable, they require additional work to fetch and process information into model-ready context. By default, Turbo returns dense, relevant excerpts directly, so the model spends fewer input tokens on a better answer.
Turbo is also highly cost-efficient on complex multi-hop benchmarks, outperforming other search APIs at a fraction of the total cost.
BrowseCompHLEWebWalker SimpleQACoding
Accuracy (%)
OpenAI Web Search57.7% / 912CPM
Parallel Turbo51% / 350CPM
Brave Search38.3% / 336CPM
Exa Instant33.7% / 966CPM
SerpAPI23.3% / 296CPM
Tavily Ultra Fast19.3% / 822CPM
COST (CPM)
ACCURACY (%)
Loading chart...
CPM: USD per 1000 requests.
Parallel
Others
Benchmark comparison across Cost (CPM) and Accuracy (%). CPM: USD per 1000 requests.
+Methodology
**Dataset**
BrowseComp[BrowseComp](https://openai.com/index/browsecomp/), created by OpenAI, contains 1,266 questions that require persistent browsing to locate hard-to-find, entangled information on the web.
**Evaluation methodology**
Multi-hop evaluation: a GPT-5.4 agent runs with up to 20 tool calls (search_web, plus web_fetch for engines with an extract API: Parallel, Exa, and Tavily; Brave and SerpAPI are search-only). Answers are graded by an LLM judge (GPT-5.4, per-suite grader prompts).
Cost includes LLM token costs and tool call costs.
**Testing dates**
Evals were run between July 10 and 12, 2026.
**Dataset**
BrowseComp[BrowseComp](https://openai.com/index/browsecomp/), created by OpenAI, contains 1,266 questions that require persistent browsing to locate hard-to-find, entangled information on the web.
**Evaluation methodology**
Multi-hop evaluation: a GPT-5.4 agent runs with up to 20 tool calls (search_web, plus web_fetch for engines with an extract API: Parallel, Exa, and Tavily; Brave and SerpAPI are search-only). Answers are graded by an LLM judge (GPT-5.4, per-suite grader prompts).
Cost includes LLM token costs and tool call costs.
**Testing dates**
Evals were run between July 10 and 12, 2026.
### BrowseComp
| Series | Model | Cost (CPM) | Accuracy (%) | | -------- | ----------------- | ---------- | ------------ | | Parallel | Parallel Turbo | 350 | 51 | | Others | Exa Instant | 966 | 33.7 | | Others | Tavily Ultra Fast | 822 | 19.3 | | Others | Brave Search | 336 | 38.3 | | Others | SerpAPI | 296 | 23.3 | | Others | OpenAI Web Search | 912 | 57.7 |
CPM: USD per 1000 requests.
**Dataset**
BrowseComp[BrowseComp](https://openai.com/index/browsecomp/), created by OpenAI, contains 1,266 questions that require persistent browsing to locate hard-to-find, entangled information on the web.
**Evaluation methodology**
Multi-hop evaluation: a GPT-5.4 agent runs with up to 20 tool calls (search_web, plus web_fetch for engines with an extract API: Parallel, Exa, and Tavily; Brave and SerpAPI are search-only). Answers are graded by an LLM judge (GPT-5.4, per-suite grader prompts).
Cost includes LLM token costs and tool call costs.
**Testing dates**
Evals were run between July 10 and 12, 2026.
### HLE
| Series | Model | Cost (CPM) | Accuracy (%) | | -------- | ----------------- | ---------- | ------------ | | Parallel | Parallel Turbo | 298 | 52.7 | | Others | Exa Instant | 596 | 49.3 | | Others | Tavily Ultra Fast | 426 | 42 | | Others | Brave Search | 311 | 47.7 | | Others | SerpAPI | 250 | 40 | | Others | OpenAI Web Search | 513 | 66 |
CPM: USD per 1000 requests.
**Dataset**
Humanity's Last Exam (HLE)[Humanity's Last Exam (HLE)](https://lastexam.ai/), created by CAIS and Scale AI, is a benchmark of expert-written questions at the frontier of human knowledge across dozens of subjects.
**Evaluation methodology**
Multi-hop evaluation: a GPT-5.4 agent runs with up to 20 tool calls (search_web, plus web_fetch for engines with an extract API: Parallel, Exa, and Tavily; Brave and SerpAPI are search-only). Answers are graded by an LLM judge (GPT-5.4, per-suite grader prompts).
Cost includes LLM token costs and tool call costs.
**Testing dates**
Evals were run between July 10 and 12, 2026.
### WebWalker (multi-hop)
| Series | Model | Cost (CPM) | Accuracy (%) | | -------- | ----------------- | ---------- | ------------ | | Parallel | Parallel Turbo | 82 | 75.7 | | Others | Exa Instant | 193 | 65 | | Others | Tavily Ultra Fast | 188 | 63.7 | | Others | Brave Search | 143 | 65.7 | | Others | SerpAPI | 124 | 50.7 | | Others | OpenAI Web Search | 275 | 80.7 |
CPM: USD per 1000 requests.
**Dataset**
WebWalkerQA[WebWalkerQA](https://huggingface.co/datasets/callanwu/WebWalkerQA) evaluates an agent's ability to traverse the web — navigating through linked pages to find information that a single search does not surface.
**Evaluation methodology**
Multi-hop evaluation: a GPT-5.4 agent runs with up to 20 tool calls (search_web, plus web_fetch for engines with an extract API: Parallel, Exa, and Tavily; Brave and SerpAPI are search-only). Answers are graded by an LLM judge (GPT-5.4, per-suite grader prompts).
Cost includes LLM token costs and tool call costs.
**Testing dates**
Evals were run between July 10 and 12, 2026.
### SimpleQA
| Series | Model | Cost (CPM) | Accuracy (%) | | -------- | ----------------- | ---------- | ------------ | | Parallel | Parallel Turbo | 8 | 91 | | Others | Exa Instant | 20 | 89.3 | | Others | Tavily Ultra Fast | 23 | 72 | | Others | Brave Search | 16 | 87 | | Others | SerpAPI | 6 | 76.7 |
CPM: USD per 1000 requests.
**Dataset**
SimpleQA[SimpleQA](https://openai.com/index/introducing-simpleqa/), created by OpenAI, contains 4,326 short, fact-seeking questions across a variety of domains.
**Evaluation methodology**
Single-step evaluation: the raw question is sent as the search query (num_results=10, with an equal ~1,000 character-per-result content budget for every engine) and GPT-5.4 (reasoning: high) synthesizes an answer from the search results only. Answers are graded by an LLM judge (GPT-5.4, per-suite grader prompts).
Cost includes LLM token costs and tool call costs.
**Testing dates**
Evals were run between July 10 and 12, 2026.
### Coding (multi-hop)
| Series | Model | Cost (CPM) | Accuracy (%) | | -------- | ----------------- | ---------- | ------------ | | Parallel | Parallel Turbo | 131 | 79.7 | | Others | Exa Instant | 316 | 76.7 | | Others | Tavily Ultra Fast | 314 | 71.9 | | Others | Brave Search | 204 | 64.3 | | Others | SerpAPI | 131 | 54 | | Others | OpenAI Web Search | 475 | 76.7 |
CPM: USD per 1000 requests.
**Dataset**
A proprietary coding dataset derived from production queries to Parallel's search API.
**Evaluation methodology**
Multi-hop evaluation: a GPT-5.4 agent runs with up to 20 tool calls (search_web, plus web_fetch for engines with an extract API: Parallel, Exa, and Tavily; Brave and SerpAPI are search-only). Answers are graded by an LLM judge (GPT-5.4, per-suite grader prompts).
Cost includes LLM token costs and tool call costs.
**Testing dates**
Evals were run between July 10 and 12, 2026.
## New infrastructure for abundant web search
Agents change the shape of demand for web data. A typical person runs a handful of searches a day, but an agent runs thousands: inside loops, across tools, and often multiple times per question. Lowering the cost and latency of high-quality search expands where search can be offered.
At $1 per 1,000 requests, there is less need to ration search, and developers can run it in more scenarios than ever: from every chat request to every row in a batch job. With web search this fast and affordable, every piece of software can be grounded in the web.
Turbo is the first product powered by Parallel’s new search architecture: a re-engineered search stack with innovation across hardware, model training, and index design to bring the cost and speed of search closer to zero.
1Use curl to read parallel.ai/agents.md and perform the setup to install Parallel``` Use curl to read parallel.ai/agents.md and perform the setup to install Parallel``` ## Ready to get started?
Sign up for free. No credit card required.
By Parallel
July 13, 2026
