Search is the new bottleneck for AI Agents
As token prices fall, web search can account for more of the cost of running an agent. A better model costs more per call, but better search can lower the cost per correct answer.
When an agent gets expensive or gives a poor answer, teams usually look at the model first. They compare providers, try a smaller model, or tune the prompt. Search deserves the same scrutiny because it determines which facts the model sees. You pay for each search call and for the tokens the model uses to read the results. To understand the cost of an answer, you need to count both.
Say your agent needs to check which of a vendor’s plans include single sign-on. If search returns a long product page that mentions the feature but leaves out the plan details, the agent spends tokens reading it and still needs another search. If the first search returns the relevant row from the vendor’s pricing table, the agent can answer correctly with fewer tool calls and fewer tokens.
Parallel Search is web search built for agents. Your agent describes what it needs in natural language and gets back ranked URLs with the most relevant excerpts from each page. The model you already use can answer from those excerpts without fetching full pages. When the first search surfaces the evidence, the agent can answer correctly with fewer tool calls and fewer tokens.
Search spend grows with every call
A few searches per task can add up to hundreds of millions of requests a year as agent usage scales to production workloads. At that scale, even a small difference in the price per search can make search the second or third largest expense in your AI stack, after inference.
Parallel Search Fast matches or exceeds the quality of other search providers at one-seventh the price. At 200 million searches a year, that is $200K instead of $1.4M. At a billion searches, it is $1M instead of $7M, a saving of $6M a year on search alone.
Annual search fees by volume
Parallel Search Fast costs one-seventh as much at every volume: $1M against $7M at 1B searches a year.
Searches per year
The $6M in savings covers search costs alone. Every search also adds text for the model to read, so fewer searches and more relevant results can reduce token spend on top of those savings.
Estimate your search bill
Monthly spend, next 36 months
Search spend by provider, next 36 months
Monthly search cost for the same workload. Hover or tap a month to read the saving, or a provider to single it out.
The same workload on other models, today
Search costs the same whichever model runs the agent, so a cheaper model makes search a larger share of the bill.
The same workload on each provider, today
Search cost per month at today’s volume.
Switching search tools cut a research agent’s token usage by 3.5×
An agent answering a research question needs to find evidence it can use to support its answer. We tested how much this kind of agent work costs when only the search tool changes.
Before, with model-native search: The agent reads through pages to find relevant information and searches again to fill gaps. Each search adds a fee and more text for the model to process as it works.
After, with Parallel: The agent gets relevant excerpts with source URLs, so it can build its answer with less reading and fewer searches.
Using our internal enterprise search benchmark, which includes questions requiring one source or information from multiple sources, we measured 3.5× fewer tokens after switching this particular research agent from model-native search to Parallel. Token usage fell from 67.1K to 19.1K per question, and total cost dropped by about two thirds with the same model and agent workflow.
What one research question costs
The same agent and model read 67.1K tokens per question with model-native search and 19.1K with Parallel Search: 3.5× fewer.
With model-native search
- User query
- Model
- Model-native search$$$
- Reasoning$$$
- Final answer
With
- User query
- Model
- Search
- Reasoning
- Final answer
Tokens per question
From prompts to production scale
Many workflows start by asking a model a question and refining the prompt based on its answer. This works well when the input already contains the information the model needs. When a question depends on something that changed yesterday, the agent workflow needs current evidence.
The first solution is often manual: search Google, open the sources, and paste the relevant passages into the model. A person manually handles the retrieval and verification. This process is a useful way to learn what the workflow needs, but every new research task requires more human effort and time.
Next, teams may connect a search tool. Some use the model provider’s built-in search. Other agent builders assemble a manual pipeline including a search API, page fetching, extraction, and validation. As volume grows, teams start asking why the agent keeps searching and how much of the retrieved text it uses.
In the next stage, teams evaluate retrieval as part of the system. They measure the tokens the agent consumes and whether it reaches a supported answer. At production scale, those measurements turn into quality targets and latency budgets for each workload.
You can adopt search designed for agents whether you’re starting with a basic prompt or already running a custom search pipeline. To optimize your model and search workflow, start by understanding the tasks your agent needs to handle.
The search maturity curve
Each stage changes how many calls an agent makes per question and what a correct answer costs.
Correct answers per dollar
What changes at each stage
13×
13× lower cost per question with Parallel, and about 15× more correct answers per dollar. Parallel answered 53.7% correctly for $15.20 per 1,000 questions. OpenAI’s built-in search answered 48.3% for $204.30.
Source: parallel.ai/benchmarks · FinSearchComp T3, Parallel Search Fast vs OpenAI native search · same agent, questions and judge · Oct 22, 2026 · cost is end to end per question; cost per correct answer derived as cost ÷ accuracy
The questions an agent answers change as a company grows. Find the questions your team asks today, and the ones it will likely ask next.
What each company stage asks of search
The questions stay simple, but the number of answers they need keeps growing.
| Ask onceA person needs one answer, now and then | Ask for every userYour product asks on each user’s behalf, all day | Ask across a listOne question, every row, run as a batch | Ask on a scheduleThe list, rechecked weekly or daily | |
|---|---|---|---|---|
| Questions | ||||
| Vendor checks | “Does Acme have a SOC 2 report?” | “Does this new vendor have a SOC 2 report?” | “Which of our 400 vendors have SOC 2?” | “Alert us when a vendor’s SOC 2 expires.” |
| Account research | “Who is the CFO at Acme?” | “Who is the CFO at this rep’s next call?” | “Who is the CFO at each of our 2,000 accounts?” | “Flag any account that names a new CFO.” |
| Pricing | “What does Acme charge per seat?” | “What does this buyer’s shortlisted tool charge?” | “What do our 50 competitors charge?” | “Tell us when a competitor changes its price.” |
| Ask onceA person needs one answer, now and then | Ask for every userYour product asks on each user’s behalf, all day | Ask across a listOne question, every row, run as a batch | Ask on a scheduleThe list, rechecked weekly or daily | |
|---|---|---|---|---|
| What changes for search | ||||
| What matters most | Getting it rightA person checks the answer | SpeedA user is waiting | ConsistencyThe same fields, each with a source | ChangeCatch what moved since the last run |
| Where the cost goes | People’s timeReading pages and pasting passages | Every requestFees and tokens on each call | Every rowA wasted search repeats down the list | Every runRereading pages that haven’t changed |
The search call changes what happens next
Search results determine what the model reads next. A relevant passage can provide enough evidence to answer, while a collection of pages requires more reading, more tokens, and often more searches.
Parallel Search returns relevant excerpts with source URLs, reducing unnecessary reading and follow-up searches. Missing evidence can lead a model to invent a plausible answer (aka Hallucinate), while irrelevant text increases token spend. By making the evidence easy to check with citations, Parallel helps teams in regulated industries verify claims and build trust in their agents’ answers.
A new filing or leadership change can make last week’s information outdated. Parallel Search provides freshness for your agents. With Parallel, you can set how recent a page must be, down to 10 minutes. If the cached copy is older than that, Parallel fetches the page in real-time. For evidence in PDFs, such as filings and court records, or websites that require JavaScript, Parallel’s Extract API converts the content into clean text the model can read.
Search has the same trade-offs as models
Developers already weigh quality, cost, and latency when they choose a model. A frontier model with high reasoning effort gives the best answers, but it costs more and responds slowly. A smaller model is cheaper and faster, and good enough for routing, classification, or simple extraction. Most production systems use more than one model, and each step gets the lowest cost model that meets its quality bar.
Search works the same way. A lookup inside a chat reply has to return in a few hundred milliseconds. A background agent checking a vendor’s compliance history can wait several seconds for deeper results. Running the deepest search on every call spends money where it isn’t needed, and running the fastest search on every call can miss evidence on hard questions. Most agents need both kinds of call, often within the same task.
Four modes, three priorities
Turbo and Fast cost $1 per 1,000 requests. Basic and Advanced cost $5.
Parallel Search has four modes, each with a different balance of quality, cost, and latency.
| Mode | When to use it | Approx. latency | Price per 1,000 requests |
|---|---|---|---|
| Turbo | When speed matters most: voice agents, chat, and high-volume lookups. | 200 ms | $1 |
| Fast | For high-quality results in interactive assistants and agents that make repeated search calls. | 700 ms | $1 |
| Basic | When the model needs longer passages from each source, with more context in a single search. | 1 second | $5 |
| Advanced | For complex questions that require finding and connecting evidence across sources, such as background research and deep reasoning. | 3 seconds | $5 |
What this looks like in production
Nooks initially used OpenAI’s built-in web search for its sales research agents, but reconsidered as research volume grew, token spend increased, and results missed recent leadership changes and security certifications. Switching to Parallel’s Search API cut web search spend by 70.5% while improving accuracy on time-sensitive facts. Read the Nooks case study
Modal built its go-to-market data pipelines in code on Parallel, using research agents to segment and enrich every prospect. Modal compared Parallel with other agentic search providers. Parallel had the highest coverage on fundraising data, at 88.9%, and cost up to 31 times less, saving $7,600 per 10,000 records. Read the Modal case study
Compare the cost of a correct answer
The next time you evaluate a model, evaluate your search too. Use real questions from your product, keep the model and agent loop fixed, and compare answer accuracy, citation support, token spend, latency, and total cost.
Evaluate search at the same time you evaluate the model. Your model determines how your agent reasons, and your choice helps determine what information your agent has to reason with. Measure cost per correct answer, including the cost of failed attempts. Then test whether better search lets a cheaper model meet your quality bar or helps your current model finish in fewer steps.
