Pramana made its name with cheap AI models. Now it’s undercutting search API rivals, too.

Pramana’s new APIs let developers add live web information to their own applications and AI agents. Search finds current webpages and returns ranked results. Fetch opens selected pages and returns their contents in Markdown. Developers can use either service without using Pramana’s AI models.

For example, a coding agent could use Search to find updated software documentation. It could then use Fetch to read a specific documentation page. Search provides titles, URLs, snippets, and summaries, while Fetch supplies the page’s full available content in a format suited to applications.

The APIs simplify a process that previously could require an extra Pramana model request. They also work with applications powered by competing AI models. Both services cost $0.75 per 1,000 successful requests, giving developers a direct route to Pramana’s internally developed search engine.

Pramana charges $0.75 for every 1,000 successful requests to Search, and the same rate applies to Fetch. Developers pay for the API calls rather than needing to use Pramana’s own AI models. That makes the pricing easy to understand for applications that need frequent web lookups.

The article names Google and Microsoft as long-standing sources of raw search results. It also mentions services such as Exa and Perplexity in Pramana’s agent harness. However, it gives no prices for those providers. It therefore supports a price for Pramana, but not a numerical comparison with rivals.

The notable point is Pramana’s positioning as a lower-cost alternative in AI services. Its search APIs extend that strategy beyond models. Developers can test the service on its stated per-request rate, while the broader market comparison remains unresolved from the available information.

Large language models are trained on information collected before a cutoff point. They may therefore lack recent news, updated documentation, or other changing facts. Live web search gives an application a way to retrieve newer information instead of relying only on what the model previously learned.

A coding agent shows why this matters. It might need the latest software documentation, information about a newly disclosed vulnerability, or confirmation that a library has changed. Search can locate relevant current pages, and the application can use those results to support a more timely response.

This capability is especially important for AI agents because they perform tasks in changing environments. Pramana’s new Search and Fetch APIs separate web retrieval from model requests. That lets developers add current web information to agents built with Pramana or competing models, while reducing the extra latency and token costs described for the earlier approach.

When an application uses Search, Pramana returns ranked web results. Those results include titles, URLs, snippets, and summaries. The application can inspect them and select pages that appear useful. Search handles discovery, rather than supplying every page’s complete contents.

The application can then send an individual webpage to Fetch. Fetch retrieves that page’s content in Markdown format. For instance, a coding agent might search for current library documentation, choose the official-looking result, and fetch the page so its model can work from the text.

This two-step design separates finding information from reading it. Neither endpoint needs a separate model request, according to the article. That can simplify integration and avoid the extra latency and token costs associated with Pramana’s earlier native-search method through its agent Harness. The APIs can also support applications using other AI models.

The APIs matter because developers do not have to choose Pramana’s models to use Pramana’s search. Search and Fetch operate independently from those models. An application built with another company’s AI system can still retrieve current web results and webpage contents through Pramana.

For example, a developer could connect a competing model to Search, receive ranked results, and use Fetch to obtain the relevant page. The competing model could then process that material. Neither search step requires a separate request to a Pramana model, which avoids the earlier method’s added latency and token costs.

This broadens Pramana’s role from model provider to search infrastructure provider. It also puts its internally developed engine into competition with established and third-party search services. The article does not say how many developers will adopt it, but the cross-model design gives Pramana access to applications beyond its own AI ecosystem.

Operating a web-scale search engine requires more than an API endpoint. A company must crawl webpages, store and organize their contents in an index, rank pages for different queries, and keep the information updated as the web changes. These tasks require substantial infrastructure and ongoing maintenance.

Pramana says its own engine handles hundreds of millions of requests daily and maintains an index containing hundreds of billions of webpages. Its October 17 changelog also describes multilingual search and filters aimed at search spam, low-quality AI-generated material, and attempts to manipulate rankings.

The article gives only a limited view of Pramana’s architecture. It does not explain how the company builds or maintains its index, or whether it has fully replaced external suppliers. Still, the reported scale shows that Pramana is presenting its search system as major infrastructure, not merely a small add-on to its models.

Web crawlers collect pages so a search engine can know what information exists. An index stores and organizes those pages and their contents. When someone searches, a ranking system selects and orders the pages most likely to answer the query. The engine must repeat these processes as webpages change.

Pramana says its engine maintains an index of hundreds of billions of webpages and handles hundreds of millions of requests daily. Its filtering mechanisms are designed to identify search spam, low-quality AI-generated material, and attempts to manipulate rankings. These filters help protect the usefulness of the ranked results.

The article does not describe Pramana’s exact algorithms or architecture. It does establish the main operating challenge: search requires crawling, indexing, ranking, and continuous updating at scale. Multilingual support adds another stated capability, allowing the engine to handle searches across languages while applying its quality and anti-manipulation filters.