Search Engine APIs and Developer Tools

By Bill Hartzer · Updated 19 August 2026

The event that reshaped this market: 11 August 2025

For years the default way to get web search results into your own software was the Bing Search APIs. They were cheap, documented, and the substrate under an enormous number of products — alternative search engines, browser extensions, research tools, and later the first wave of AI answer products.

Microsoft retired them on 11 August 2025. Not deprecated, not price-adjusted: retired. Microsoft's lifecycle announcement is unambiguous — the APIs "will be retired on August 11, 2025", "any existing instances of Bing Search APIs will be decommissioned completely", and the product is "no longer available to be used or new customer signup". Customers were directed to Grounding with Bing Search inside Azure AI Agents.

That replacement is not the same product. The retired APIs returned a ranked list of web results you could do anything with. Grounding with Bing Search is a tool an Azure AI Agent calls so a large language model can incorporate live web data when generating a response — a component of an LLM pipeline, not a search endpoint. It does not return ranked links in the same form, it is metered inside Azure rather than sold standalone, and it carries display and use terms attached to the agent context.

Two consequences followed. Every product quietly reselling Bing had to find an index elsewhere or stop; and the handful of organisations that own a web-scale index and will sell access to it became considerably more valuable. That is why several entries below launched or expanded their API businesses within eighteen months of that date.

Brave Search API

What is behind it: Brave's own index, built by its own crawler — Brave describes tens of billions of pages with over a hundred million updates a day, supplemented by the opt-in Web Discovery Project. Brave removed the last of its Bing dependencies from web results in April 2023, so this does not resolve to Google or Microsoft underneath.

What it returns: a broad endpoint set rather than a single search call — web results, an LLM-context endpoint shaped for retrieval-augmented generation, a summarised Answers endpoint with citations, plus image, video, news, suggestion and spellcheck endpoints.

How the pricing is structured: two self-serve plan families and a negotiated enterprise tier. The search family is metered per block of requests. The answers family charges per request and meters input and output tokens on top, because generation cost scales with text volume rather than call count — a token-metered endpoint behaves nothing like a request-metered one under load. A recurring free credit allowance applies automatically.

Brave markets the API explicitly on independence: AI and agent developers needing grounding data have very few options that are not Google or Microsoft. That pitch got considerably stronger on 11 August 2025.

Google Custom Search JSON API — and its end date

What it is: the long-standing way to get Google-quality results programmatically, returning JSON to the OpenSearch 1.1 specification over a Programmable Search Engine you configure first.

The limitation people miss: it searches the sites you configured, not the open web. Google's documentation is explicit that full web search requires a different arrangement. Assuming you are buying "the Google index in JSON" and discovering you bought site search over your own domain list is the most common mistake in this area.

How the pricing is structured: a small free daily query allowance, then per-thousand-query billing above it, subject to a hard daily ceiling. It is quota-shaped rather than capacity-shaped — you cannot buy your way past the daily cap on the self-serve tier.

The material fact, checked 19 August 2026: the Custom Search JSON API is closed to new customers, and Google's own overview page states that existing customers must move to an alternative by 1 January 2027. Google points those needing search over up to fifty domains toward Vertex AI Search, and invites contact for full web search. Anyone planning a build on this API in 2026 is planning a migration.

The pattern is worth naming: both Microsoft and Google have now withdrawn a general-purpose self-serve web search API and pointed customers at a grounding product inside a managed AI platform. That is a repricing of web index access from a commodity into an AI-platform feature.

You.com — a consumer engine that became an API company

You.com is the clearest case study in this shift. It launched as a consumer AI search engine, could not win attention against ChatGPT and Google, and repositioned entirely. Checked 19 August 2026, you.com presents no consumer search box; the headline is "The Leading Web Search APIs for AI".

What is behind it: its own crawler, YouBot, and its own index; You.com documents the crawler's IP range and reverse-DNS verification. In its consumer era it was widely understood to blend third-party feeds, so treat the current independence claim as documented but unaudited.

What it returns: a stack rather than a single endpoint — a Web Search API for results, a Contents API returning clean extracted page text across multiple URLs, an Answer API for grounded cited answers, a Research API, a Finance Research API, an MCP server for agent integration, and agent skills.

How the pricing is structured: per-call billing on each API, sold direct and through AWS Marketplace, with a free developer credit allowance used for acquisition. This is infrastructure pricing to software companies, not consumer monetisation.

The commercial validation is real: a September 2025 Series C valued the company in the billions on the API business, not the search engine. Named customers as of October 2025 included Databricks, Harvey AI, Windsurf and — notably — DuckDuckGo, meaning one engine here draws on another's index.

Perplexity: Sonar, the Search API, and a migration in progress

Perplexity sells two structurally different things, and conflating them causes confusion.

Sonar is the grounded-generation line: you send a question, it retrieves and returns a written answer with citations. The Sonar model family spans a base model, a Pro tier, a reasoning variant and a deep-research variant, differing in how much retrieval and reasoning each performs per call.

The Search API, launched 25 September 2025, is the opposite shape: raw ranked results from Perplexity's own index, with filtering and no generation. It can sell this only because it owns an index — its engineering write-up of the same date describes an exabyte-scale crawling apparatus tracking over 200 billion unique URLs, which is also why the Bing retirement six weeks earlier did not affect it.

What is changing, checked 19 August 2026: Perplexity's documentation states that Sonar Chat Completions is now the Agent API and displays a migration deadline of 27 September 2026. The Agent API changes the request and response shape — an input parameter returning a typed output array with per-step detail, rather than a messages array returning choices — and adds multi-step agentic loops, code sandboxes, MCP servers, and model fallback across vendors. Anyone running Sonar Chat Completions in production has a dated migration ahead.

How the pricing is structured: generation endpoints meter tokens and requests together, so cost scales with how much retrieval and reasoning a call performs, not just with call count. The Search API is metered on requests.

The smaller independent indexes

Three more organisations sell access to an index they actually own, and they matter out of proportion to their size because there are so few of them.

Mojeek — a metered Web Search API over its own British-built index, plus a Site Search API for site-scoped search. Its index has scarcity value precisely because it is not Google, Bing, Yandex or Baidu underneath.
Kagi — a Search API, opened in preview to all users in 2026, plus AI endpoints: FastGPT for fast grounded answers with citations, and a Universal Summarizer, backed by Kagi's own Teclis and TinyGem indexes. Note the caveat that applies to Kagi generally: much of a Kagi results page comes from anonymised calls into third-party commercial indexes, so an API over Kagi is not an API over a wholly owned index.
Marginalia Search — a keyed API over a small, deliberately non-commercial index under its own licence terms, and one of SearXNG's pluggable engines. Not a general-purpose web API, and it does not pretend to be: it is the right tool only when you want the text-heavy corner of the web mainstream ranking buries.

Also worth watching: European Search Perspective, the joint venture co-owned by Ecosia and Qwant, whose Staan index has been serving queries since August 2025 and is explicitly pitched at AI companies and third-party search products wanting a European, non-US source of web results.

A non-API option: SearXNG is free self-hostable metasearch software that forwards queries to other engines and merges the results. It is neither an index nor a commercial API — fine for low-volume internal use, bad at any scale, because it depends on upstream engines that never agreed to supply you.

Five questions to ask before committing to a search API

Whose index is actually underneath? If the supplier does not run a crawler, you are exposed to their supplier's decisions. The Bing retirement removed a substrate under dozens of products in a single day.
Is it a search endpoint or a grounding tool? Ranked links you can process yourself, and a web-grounding tool for an LLM inside a managed platform, are different products with different terms and different failure modes.
How is it metered? Per-request pricing is predictable; token-metered generation endpoints scale with output volume and can behave very differently under load. Several of the APIs here charge both ways on the same call.
What are the caching, storage and display terms? Search APIs commonly restrict how long results may be stored, whether they may be re-displayed, and whether they may train a model. These clauses are where an otherwise cheap API becomes unusable.
What is the announced end date? In this market that is now a real question rather than a paranoid one. Two of the largest general-purpose search APIs have been withdrawn within eighteen months of each other.

Microsoft's own lifecycle announcement remains the primary record of the August 2025 retirement, and Google's Custom Search JSON API overview carries the closed-to-new-customers notice and the 1 January 2027 transition date.

When this page was last checked

Every claim on this page was verified against the provider's own documentation on 19 August 2026. The dated facts are: the Bing Search APIs retired 11 August 2025, replaced by Grounding with Bing Search in Azure AI Agents; the Google Custom Search JSON API closed to new customers with a stated transition deadline of 1 January 2027; Perplexity's Search API launched 25 September 2025; and Perplexity's documentation showed a Sonar-to-Agent-API migration deadline of 27 September 2026.

No prices appear anywhere on this page, deliberately. Every provider listed has changed its rate card at least once and several have restructured tiers entirely; a stated price is the fastest-rotting fact in this subject. Pricing is described structurally — per-request, per-thousand-requests, token-metered, quota-capped, subscription or negotiated — because the shape of the meter is stable in a way the numbers are not.

Frequently asked questions

Is the Bing Search API still available?

No. Microsoft retired all Bing Search APIs on 11 August 2025, decommissioning existing instances completely and closing new signups. The stated replacement is Grounding with Bing Search inside Azure AI Agents, which is a web-grounding tool for a large language model rather than a raw search endpoint. Products that had been reselling Bing results through the public API had to migrate or stop.

What replaced the Bing Search API?

Grounding with Bing Search, a tool available to Azure AI Agents that lets an LLM incorporate live public web data when generating a response. It is not a drop-in replacement: it returns grounded generation inside an agent pipeline rather than a processable list of ranked links, it is metered inside Azure, and it carries use and display terms attached to that context.

Can I use the Google Custom Search JSON API?

Not as a new customer. As checked on 19 August 2026, Google's overview page states the Custom Search JSON API is closed to new customers and that existing customers must transition to an alternative by 1 January 2027. It also only searches the sites configured in your Programmable Search Engine, not the open web — a limitation many developers discover late.

Which search APIs run their own web index?

Brave Search, You.com, Perplexity, Mojeek and Marginalia all sell access to indexes built by their own crawlers. Kagi sells an API but sources much of a results page from third-party indexes, so it is a partial case. This matters because an API over a borrowed index carries the borrowing arrangement's risk, as the Bing retirement demonstrated.

How is search API pricing usually structured?

Three shapes recur. Raw search endpoints are metered per block of requests. Generated-answer endpoints often charge per request and additionally meter input and output tokens, so cost tracks text volume rather than call count. Quota-shaped APIs give a free daily allowance then bill per thousand queries under a hard ceiling. Free recurring credit allowances are common for acquisition.

What happened to You.com's search engine?

It was discontinued as a consumer product. You.com now sells its index and derived products as APIs to other software companies — a Web Search API, a Contents API for clean page text, Answer, Research and Finance Research APIs, an MCP server and agent skills. Its homepage, checked 19 August 2026, has no consumer search box. It failed as an engine and succeeded as infrastructure.

Is Perplexity's Sonar API being discontinued?

It is being consolidated. As checked on 19 August 2026, Perplexity's documentation states that Sonar Chat Completions is now the Agent API and displays a migration deadline of 27 September 2026. The Agent API uses a different request and response shape and adds multi-step agentic loops, code sandboxes, MCP servers and model fallback. Perplexity advises using it for all new projects.

What should I check in a search API's terms?

Caching, storage and display restrictions, and whether results may be used to train a model. Search APIs routinely limit how long results can be retained, whether they can be re-displayed to end users, and what derived use is permitted. These clauses, rather than the rate card, are what most often make an otherwise suitable API unusable for a particular product.

Search engine APIs and developer tools