Agentic web search splits into three layers: $1–$10 retrieval, $10–$45 model-native grounding, and research agents whose cost varies 25-fold
Asked:
“Who provides agentic websearch capabilities, how do they compare on performance and cost, and how do their architectures differ?”
This scan, as of 7 October 2026, covers 19 representative provider rows — one row per source, across 18 independent hosts — plus a 150-row benchmark evidence pool. It is representative, not exhaustive: more providers exist. Prices are current list prices in US dollars per 1,000 calls where the vendor publishes them; credit-priced and token-metered offerings are left unresolved rather than converted.
The landscape: architecture layer vs cost per 1,000 calls
Each dot or bar is one product, placed on a log cost scale. Colour is what you get back, not who sells it. Hover or tap a mark for the pricing detail and source. Google’s three rows (Gemini Developer API, Google AI, Google Cloud) appear as one family of grounding variants.
Not placeable on this scale:
Ordinary retrieval is cheap and converging: Parallel and Perplexity fast modes start near $1/1K, You.com and standard tiers sit at $5/1K, Exa at $7/1K, Anthropic at $10/1K plus tokens — a 10× spread, not an order-of-magnitude war. parallel.ai, exa.ai
Model-native grounding couples search spend to token spend: OpenAI charges $10/1K (o-series) or $25/1K (GPT-4o/4.1) plus model tokens; Gemini 3 grounding is $14/1K after 5,000 free prompts a month, Gemini 2.5 is $35/1K. community.openai.com, ai.google.dev
Deep research is where cost variance explodes: Parallel Responses spans $10/$50/$250 per 1K by reasoning effort, Linkup Research runs $0.25–$2.50 per single request, and OpenAI, Gemini, and Perplexity deep research meter tokens, so actual cost depends on loop length. parallel.ai, docs.linkup.so
Architectural ownership is the real differentiator: Exa and Parallel run their own web-scale indexes with semantic ranking and compressed excerpts; Anthropic can have Claude write code to filter results before they enter context; Firecrawl and Jina own the browsing and extraction layer for dynamic pages. exa.ai, platform.anthropic.com, firecrawl.dev
Benchmarks: two tests, two layers — do not compare across them
Evaluations test different layers and workloads, so there is no credible universal winner. The two panels below use different metrics on different dates and are not numerically comparable with each other, nor with model-agent scores such as OpenAI’s reported 26.6% on Humanity’s Last Exam at its deep-research launch.
Artificial Analysis Search Index — overall score (0–100)
Eight-API retrieval test — Agent Score, with latency
Evaluated December 2025 · llms.blog (top 4 of 8 shown)
Caveats: the September 2026 index is widely recapped by vendors (Parallel, TinyFish, Octen) with inconsistent cost units across recaps; TinyFish reports 71.2 overall at $0.0345 and 59.4 s per task while noting weaker DeepSearchQA performance. Vendor claims such as Exa’s retrieval accuracy are architecture evidence, not neutral head-to-head proof. Missing scores are omitted, never ranked as zero.
Provider comparison
All 19 source rows. Prices are vendor list strings; blank cells mean the vendor publishes none.
Provider
Product
Search / entry price
Research price
Free tier
Source
Which to pick
Cheapest high-volume retrievalParallel turbo/fast or Perplexity fast modes at ~$1/1K; Serpent lists $0.03–$0.60/1K at scale tiers for plain SERP JSON.
Best semantic / content retrievalExa ($7/1K, own neural index, content and highlights included) or Parallel Search (objective-ranked compressed excerpts from a proprietary index).
Easiest model-integrated cited answersAnthropic’s web search tool, OpenAI’s web search, or Gemini grounding — one flag, the model decides when to search and returns inline citations; search spend rides on token spend.
Best multi-hop structured researchParallel Task (per-field provenance, effort tiers) or Linkup deep mode (up to 10 retrieval iterations, schema-constrained JSON); OpenAI/Gemini/Perplexity deep research when you accept token-metered cost.
Best dynamic-page extractionFirecrawl — JavaScript rendering, browser interaction, crawl, structured markdown; Jina Reader for simple URL-to-LLM content.
Unresolved pricing — budget carefullyTavily (credit-priced), Firecrawl credits, and all token-metered deep research: cost depends on cache hits, page counts, or loop length.
Method: 19 representative provider rows scanned as of 2026-10-07, one source URL per row across 18 independent hosts (max share of any URL 5.3%), holding list price strings, free tiers, and architecture notes; plus a 150-row benchmark evidence pool from which only clearly named benchmarks with plausible units were kept (Artificial Analysis Search Index, Sep 2026; eight-API Agent Score retrieval test, Dec 2025). Costs are vendor list prices per 1,000 calls in USD; credit-priced and token-metered products are left unresolved. Duplicate benchmark recaps, rows with suspect cost units, and four of the eight Dec-2025 APIs lacking scores in the pool were cut for space. Jina Reader is described from the brief’s architecture context and has no priced row in this scan.