Parallel Fast Beats Claude's Search While Cutting Costs by a Third

AI search infrastructure company Parallel, led by former Twitter CEO Parag Agrawal and also responsible for @p0, took the top two positions on Vals AI’s independent Web Search Index. With Claude Opus 5.5 held constant, Parallel Fast scored 57.5%, Parallel Advanced scored 54.1%, and Claude’s native web search scored 53.8%.

For agent developers, the retrieval swap raised the displayed score by 3.7 percentage points while reducing reported cost and latency by about one-third. The result shows how much an agent’s search provider can affect end-to-end performance without changes to the model or prompt.

A benchmark that forces retrieval

Web search evaluations can overstate retrieval quality when a model already knows the answers. Vals cites a 2026 study in which a frontier model answered 44.5% of BrowseComp questions without search, suggesting that training exposure or existing model knowledge can obscure the search tool’s contribution.

The Vals benchmark uses expert-written tasks from FAB v2, the Finance Agent Benchmark, and the Legal Research Benchmark. Vals removes direct domain tools such as SEC EDGAR integrations, requiring agents to gather external evidence through general web search.

Vals also runs a no-search control. Agents score 2.9% on legal tasks and 7.4% on finance tasks without retrieval, while search-equipped configurations reach roughly 30% to 50%. The model and agent harness remain fixed across runs, with only the search provider changing. Within that setup, score differences reflect the retrieval layer and how effectively the agent can use its results.

Fast wins on accuracy

The displayed scores put Fast 3.4 percentage points ahead of Advanced and 3.7 points ahead of native search. Advanced cost six cents less per completed task, despite occupying a more expensive API tier, because task-level cost depends on the number of searches and model turns used during an agent run.

Vals reports full task latency, which differs from the latency of a single search request. An agent can spend several minutes making repeated searches, reading results, and generating intermediate reasoning even when each Fast API call returns in under a second.

Parallel lists Turbo and Fast at $1 per 1,000 requests. Turbo targets a p50 latency of about 200 milliseconds, meaning half of requests complete within that time. Fast targets roughly 700 milliseconds and allocates more time to result quality. Those request-level prices and latency targets are separate from the benchmark’s end-to-end task totals.

Search with an objective

Parallel’s API lets an agent describe the information it needs through an objective field and request a bounded number of results. The response is designed for direct use in a model’s context window, reducing the work required to turn ranked web pages into usable evidence.

The parallel-web SDK is available through PyPI and npm. A Python request requires an API key in the PARALLEL_API_KEY environment variable:

from parallel import Parallel

client = Parallel()
results = client.search.create(
    objective="Columbus-based corporate law firms specializing in disability care",
    max_results=10,
)

Accounts receive $5 in monthly credits, covering up to 5,000 Turbo or Fast searches at the listed rate. Developers can also pass the API key to Parallel’s Model Context Protocol server as a Bearer token, which raises rate limits and enables the web_research tool.

Choose for the agent loop

Because the benchmark uses legal and finance questions, its strongest evidence applies to case research, regulatory lookups, SEC filing analysis, and corporate due diligence. Customer support, factual question answering, consumer chat, and voice agents have different latency and coverage requirements and need workload-specific evaluation.

• Turbo, about 200ms p50: suited to voice interfaces, consumer chat, and frequent grounding calls where response time dominates.
• Fast, about 700ms: a strong candidate for general agent loops and the highest-scoring tier in this benchmark.
• Basic and Advanced: intended for deeper research passes where richer per-call retrieval can justify additional latency or price.

How far the result travels

The ranking covers two professional domains, one model, and one agent harness. Fast’s 3.4-point lead over Advanced is meaningful within this run, while Advanced’s 0.3-point lead over native search is small enough to warrant checking sample size, variance, and future reruns before drawing a broad conclusion.

For production systems, search provider and tier now belong in the evaluation matrix alongside model, prompt, and tool design. A controlled provider swap moved the benchmark score by several points and cut task cost and latency by about one-third, giving developers a concrete reason to test retrieval independently from the model.