{"slug": "deepseek-search-and-fetch-api-vs-firecrawl-exa-and-tavily-the-web-access-slot", "title": "DeepSeek Search and Fetch API vs Firecrawl, Exa and Tavily: the web access slot priced per 1,000", "summary": "DeepSeek's web access API, launched 17 October 2026, offers the cheapest search among six compared providers at $0.75 per 1,000 successful requests, according to a comparison checked against vendor pages on 19 October 2026. Firecrawl, Exa, Tavily, Parallel and the Brave Search API charge $1 to $8 per 1,000 searches, and only Firecrawl's yearly Growth and Scale plans undercut DeepSeek on fetch at $0.67 and $0.60 per 1,000 pages if every credit is used. The comparison notes DeepSeek's guide documents no robots.txt, user-agent, egress or query-retention policy, while its privacy policy last updated 10 February 2026 states personal data is stored in the People's Republic of China and inputs may be used for training with an email opt-out.", "body_md": "# DeepSeek Search and Fetch API vs Firecrawl, Exa and Tavily: the web access slot priced per 1,000\n\nAs of 19 October 2026, [DeepSeek](https://stackness.dev/tools/deepseek-search-fetch-api)'s web access API, launched on 17 October, has the cheapest search of the six compared here, at $0.75 per 1,000 successful requests. [Firecrawl](https://stackness.dev/tools/firecrawl), [Exa](https://stackness.dev/tools/exa-ai), [Tavily](https://stackness.dev/tools/tavily), [Parallel](https://stackness.dev/tools/parallel-web) and the [Brave Search API](https://stackness.dev/tools/brave-search-api) charge $1 to $8 per 1,000 searches. For most agents the pick is a search API that returns page content in the same call, which Exa includes at no extra charge for up to 10 results. DeepSeek is the bulk option for data you are content to process in China.\n\n## What do the DeepSeek Search and Fetch APIs return?\n\nThe DeepSeek Search and Fetch APIs are two JSON endpoints that use the same key and balance as DeepSeek's chat API. Search takes a query of up to 128 characters and returns up to 50 results, each with a title, URL, a short snippet, a 600-800 character summary \"intended to be read by a model\" and, on request, the page content. Fetch takes one URL and returns the page as Markdown.\n\nThe [guide](https://api-docs.deepseek.com/guides/search/) lists search operators (`site:`, `-site:`, quoted phrases, `lang:`, `after:` and `before:`) and two scores per result: `rerank_score` from 0 to 3 and `authority` from 0 to 4. Billing counts only a 200 response with a result. Empty results, failed fetches and timeouts are free, and `count` does not change the price, so 50 results with content cost one request.\n\nOne detail changed in the first days. The rate limits were 5 queries a second for each endpoint in every snapshot of the guide up to 18 October; on 19 October the page says 100 for Search and 25 for Fetch. Calls made through [DeepSeek Harness](https://stackness.dev/tools/deepseek-harness), [Claude Code](https://stackness.dev/tools/claude-code) or [Codex](https://stackness.dev/tools/codex-cli) as server-side tools on DeepSeek's API are free until 1 November 2026 at 00:00 Beijing time (16:00 UTC on 31 October, per the [release note](https://api-docs.deepseek.com/news/news261017)). The [Claude Code provider post](https://stackness.dev/blog/run-claude-code-with-a-different-model-provider-env-vars-what-breaks) covers pointing Claude Code at DeepSeek.\n\nThe guide says nothing about robots.txt, the user agent, egress addresses or how long queries are kept. DeepSeek's [privacy policy](https://cdn.deepseek.com/policies/en-US/deepseek-privacy-policy.html), last updated 10 February 2026, says it stores personal data \"in People's Republic of China\" and may use inputs to train its models, with an opt-out by email, depending on where you live. No zero data retention option is documented.\n\n## How I picked the six\n\nI kept every provider with a public per-request price for search, fetch or both, and checked each price on the vendor's own page on 19 October 2026. Where a vendor sells credits, I divided the plan price by its credits; those rows say which plan.\n\nLeft out, and why:\n\n- **Jina Reader** bills per output token, and I could not find its dollar rate on an official page.\n- **Perplexity** tops the[Artificial Analysis Search Index](https://artificialanalysis.ai/agents/search-api) at 80, but its API bills differently enough to need its own comparison.\n- **SerpAPI, Serper, Kagi and Linkup** were quoted by Hacker News commenters; I did not check them against their own pages.\n- **Self-hosted SearXNG** costs nothing per request, but commenters report upstream engines rate-limit it quickly.\n\n## Six web access APIs an agent can call\n\n### DeepSeek Search and Fetch\n\nClosed service · $0.75 per 1,000 successful searches, $0.75 per 1,000 successful fetches · Markdown fetch, no crawl · launched 17 October 2026\n\nThe cheapest search by a factor of 1.3 against the next option, Parallel's $1; for fetch, only Firecrawl's yearly Growth and Scale plans go lower per page, at $0.67 and $0.60 per 1,000 if every credit is used. The [Artificial Analysis index](https://artificialanalysis.ai/agents/search-api) added DeepSeek Search on 19 October at 69, below Exa auto (74), Firecrawl (73) and Parallel Basic (73) and above Tavily basic (66). On [Hacker News](https://news.ycombinator.com/item?id=50138531), one commenter who ran it on their own fetch benchmark wrote that \"its search and fetch aren't very good (is quite cheap tho)\", and a commenter from Keenable, which sells a competing search API, said it scored better than Jina, Firecrawl and Tavily on its NEEDLE benchmark. Neither commenter posted numbers. Pick it for high-volume, non-sensitive lookups where price decides.\n\n### Firecrawl\n\nAGPL-3.0, self-hostable without its anti-bot engine · 2 credits per 10 search results, 1 credit per scraped page · Markdown, HTML, JSON, screenshots · tag v2.11.542, 19 October 2026\n\nFirecrawl is a scraper and crawler first, with search on top. Per 1,000, a 10-result search costs $7.60 on the $19 Hobby plan, $1.66 on the $83 Standard plan billed yearly and $1.20 on the $599 Scale plan; a scraped page costs $3.80, $0.83 and $0.60 ([pricing](https://www.firecrawl.dev/pricing), [billing](https://docs.firecrawl.dev/billing)). It renders JavaScript, runs page actions, crawls whole sites and parses PDF, DOCX, XLSX and PPTX. A returned document is billed even when the page answered 403 or 404. Pick it when the agent needs whole sites or documents, not only answers.\n\n### Exa\n\nClosed service, MIT SDK · $7 per 1,000 auto searches, $4 instant, contents included up to 10 results · Markdown text, highlights, summaries · exa-py v2.25.0, 1 October 2026\n\nExa returns page text with search results at \"no extra charge up to 10 results per search\" ([contents docs](https://exa.ai/docs/reference/contents-retrieval)), so one $7 call replaces a search plus ten fetches. Standalone contents cost $1 per 1,000 pages per content type ([pricing](https://exa.ai/pricing)). It handles JavaScript pages and PDFs, and its auto mode scored 74 on the Artificial Analysis index. Pick it as the default for agents that read what they find.\n\n### Tavily\n\nClosed service, MIT SDK · $8 per 1,000 basic searches pay-as-you-go, $5 on Growth · extract $1.60 per 1,000 URLs · Markdown or text\n\nTavily bills credits at $0.008 each pay-as-you-go: 1 for a basic search, 2 for advanced, 1 per 5 extracted URLs ([credits](https://docs.tavily.com/documentation/api-credits)). Failed extractions are free. Its docs say Extract handles JavaScript-rendered pages but say nothing about PDFs or how it treats robots.txt, and its basic search scored 66 on the Artificial Analysis index. Pick it if your framework already ships a Tavily tool and the volume is small.\n\n### Parallel\n\nClosed service · $1 per 1,000 Turbo or Fast searches, $5 Basic or Advanced · extract $1 per 1,000 URLs · Markdown excerpts or full pages\n\nParallel has the cheapest search after DeepSeek, at $1 per 1,000 fast searches and $1 per 1,000 extracted URLs, with up to 5,000 free requests a month ([pricing](https://parallel.ai/pricing)). Its Advanced mode scored 75 on the Artificial Analysis index and Basic 73. Parallel publishes comparison pages against Tavily and Exa, so read its own claims as a vendor's. Pick it for a cheap search-plus-extract pair from one vendor.\n\n### Brave Search API\n\nClosed service · $5 per 1,000 searches, $5 of free credit a month · snippets, no fetch endpoint · 50 queries a second\n\nBrave runs its own index and returns snippets, with up to five extra excerpts per result ([Brave Search API](https://brave.com/search/api/)). It has no endpoint to fetch an arbitrary URL, so it covers half the slot. Its LLM Context mode scored 75 on the Artificial Analysis index. Pick it for search from an independent index, paired with any fetch API.\n\n## The six side by side\n\nPrices per 1,000 requests from each vendor's page on 19 October 2026:\n\n| Option | Search | Fetch or extract | Billed on | Free tier | Crawl | Licence | \n|---|---|---|---|---|---|---|\n| DeepSeek | $0.75 | $0.75 | Successful requests only | None (free via server tools until 1 Nov) | No | Closed | \n| Firecrawl | $1.20-$7.60 | $0.60-$3.80 | Credits, blocked pages included | 1,000 credits a month | Yes | AGPL-3.0 | \n| Exa | $4-$7, contents included | $1 | Requests | $10 a month | No | Closed | \n| Tavily | $5-$8 | $1.00-$1.60 | Credits, failed extracts free | 1,000 credits a month | Yes | Closed | \n| Parallel | $1-$5 | $1 | Requests | 5,000 requests a month | No | Closed | \n| Brave | $5 | none | Requests | $5 a month | No | Closed | \n\nOne search plus three fetches, a common agent step, costs $0.003 on DeepSeek, $0.004 on Parallel Fast plus Extract, $0.007 on Exa (one auto search with contents) and $0.019 on Firecrawl Hobby. Among the [AI tools developers wire into their agents](https://stackness.dev/categories/ai-tools), none of these six is on a public Stackness stack without bots as of 19 October 2026 ([data sources](https://stackness.dev/about/data-sources)).\n\n## A search API sits between the agent and the scraper\n\nA web access API takes the slot between the agent's model and a full scraper. Search finds URLs and a summary; fetch turns one URL into Markdown; a scraper such as Firecrawl's crawl mode walks a whole site with actions and proxies. A vector index, like the ones in the [Postgres vector post](https://stackness.dev/blog/pg-vectorsearch-vs-pgvector-prism-vs-hnsw-and-the-postgres-vector-slot), holds your own corpus and costs nothing per query.\n\nMost agents need the first two and never the third. DeepSeek's own harness shows why fetch belongs in a hosted service: Firecrawl's [plugin for DeepSeek Harness](https://github.com/firecrawl/dsh-firecrawl) notes that the base harness ships `web_fetch` disabled, because its only local backend \"defers SSRF protection and the model would choose the request target\". The [harness comparison](https://stackness.dev/blog/deepseek-harness-vs-claude-code-which-harnesses-lock-you-to-one-model-and-what-breaks-when-you-switc) covers the rest of DeepSeek's harness.\n\n## Blocked pages cost money on some APIs and nothing on others\n\nSites blocking agents is a cost question as well as a coverage one. DeepSeek returns a 404 for \"a URL blocked by site policy\" and does not bill it. Tavily does not bill failed extractions. Firecrawl bills any page it returns, 403 and 404 included, and escalates to an enhanced proxy automatically at no extra credit. Exa reports a status per URL.\n\nBlocking is spreading. Cloudflare said in a [July 2026 announcement](https://www.cloudflare.com/en-au/press/press-releases/2026/cloudflare-allows-the-agentic-internet-to-flourish-with-a-simple-philosophy-your-content-your-rules/) that from 15 September its new customers and free customers who kept the defaults get defaults that \"allow for search but block training and agent use for pages with ads\". A fetch API acting for a user falls in its Agent category. [The Verge](https://www.theverge.com/ai-artificial-intelligence/1013195/ai-agents-blocked-web) reported on 17 October that Yelp lets in only agents that pay for its data licensing, Delta stops agents booking tickets, and JPMorgan Chase and Charles Schwab block agents from customer accounts.\n\nSite owners have reasons. Konstantin Ryabitsev [counted](https://people.kernel.org/monsieuricon/creepy-crawlies) about 6.1 million daily requests for git.kernel.org commit pages and estimated that at most 1.9% could plausibly come from a person. Providers differ in how they identify themselves: Exa names its index crawler `ExaSearchBot` and its on-request fetcher `ExaUser` ([crawler page](https://crawler.exa.ai)), Firecrawl honours a `FirecrawlAgent` robots.txt token when crawling, and DeepSeek documents nothing.\n\n## Do you still need Firecrawl or MarkItDown for documents?\n\nYou still need a converter for documents if your agent reads PDFs, Word or spreadsheet files. DeepSeek Fetch does not document document support and lists \"unsupported content type\" among its 404 reasons. Exa handles PDFs. Firecrawl converts PDF, DOCX, XLSX, PPTX, EPUB and more, from a URL or an uploaded file ([document parsing](https://docs.firecrawl.dev/features/document-parsing)).\n\nFor files you already have, a local converter costs nothing per request. [MarkItDown](https://stackness.dev/tools/markitdown), MIT-licensed from Microsoft, turns PDF, Office files, HTML and more into Markdown; v0.1.8 shipped on 21 September 2026. Its README warns that OCR of embedded images is \"silently skipped\" without an LLM client. [RAGFlow](https://stackness.dev/tools/ragflow), Apache-2.0, parses layouts and tables for a full retrieval pipeline and asks for 4 cores and 16 GB of RAM to start.\n\n## Firecrawl, Exa and Tavily document a limit per key\n\nA per-request web tool has no natural ceiling: an agent in a loop can search thousands of times. Firecrawl sets a credit limit per API key per day, week or month, plus a monthly pay-as-you-go cap. Exa sets a budget per key and returns 402 when it is reached. DeepSeek runs on a prepaid balance and stops at zero.\n\nTavily returns HTTP 432 when a key's usage limit or the plan's credit limit is reached, and 433 when the pay-as-you-go limit set in its dashboard is reached. Parallel's monthly spend limit, per organization or per app, only emails admins and never blocks requests, and I found no spend cap in Brave's docs. For the rest, cap the loop in the harness: a maximum number of tool calls per task, or a hook that counts calls. The [subagent cost post](https://stackness.dev/blog/coding-agent-spawned-subagents-runaway-cost-causes-tests-and-caps) shows the hook pattern.\n\n## Which one for which agent\n\n- **A research or coding agent that reads what it finds:** Exa, one call per search with contents.\n- **High volume, public data, price first:** DeepSeek, after checking its results on your own queries.\n- **Cheapest from a vendor outside China:** Parallel Fast search with Extract.\n- **Whole sites, JavaScript actions or documents:** Firecrawl.\n- **Search from an independent index:** Brave, plus any fetch API.\n- **An existing framework with a Tavily tool and small volume:** Tavily.\n- **Your own files:** MarkItDown locally, no API at all.\n\n## New posts by email\n\nAbout one email a week with what we published. No account needed.\n\n## Which of these do you use?\n\nNobody on Stackness runs DeepSeek Search and Fetch API yet. Tick what you use and be the first.\n\n[Play Lonny Eagle Amateur Skater 4A pixel art skateboarding game. Signing up opens it straight away.Sign up to play](https://stackness.dev/games/lonny-eagle)", "url": "https://wpnews.pro/news/deepseek-search-and-fetch-api-vs-firecrawl-exa-and-tavily-the-web-access-slot", "canonical_source": "https://stackness.dev/blog/deepseek-search-and-fetch-api-vs-firecrawl-exa-tavily-price-per-1000", "published_at": "2026-10-19 08:06:34+00:00", "updated_at": "2026-10-19 09:16:59.211873+00:00", "lang": "en", "topics": ["ai-tools", "ai-agents", "ai-search", "ai-crawlers"], "entities": ["DeepSeek", "Firecrawl", "Exa", "Tavily", "Parallel", "Brave Search API", "Claude Code", "Codex"], "also_reported_by": [], "alternates": {"html": "https://wpnews.pro/news/deepseek-search-and-fetch-api-vs-firecrawl-exa-and-tavily-the-web-access-slot", "markdown": "https://wpnews.pro/news/deepseek-search-and-fetch-api-vs-firecrawl-exa-and-tavily-the-web-access-slot.md", "text": "https://wpnews.pro/news/deepseek-search-and-fetch-api-vs-firecrawl-exa-and-tavily-the-web-access-slot.txt", "jsonld": "https://wpnews.pro/news/deepseek-search-and-fetch-api-vs-firecrawl-exa-and-tavily-the-web-access-slot.jsonld"}}