Sofya is now on NanoGPT: search, fetch, extract, and deep research

We added Sofya to NanoGPT.

You can use Sofya as a web search provider for chat, or use Sofya Research as its own model when you want a longer report with sources and sub-queries.

What Sofya is

Sofya is a web toolkit for AI agents: search, fetch, extract, and research behind one API.

• Search: web search that returns extracted page content, not just snippets.
• Fetch: URL to clean markdown, including webpages, PDFs, DOCX files, and other document formats.
• Extract: AI-powered extraction from a page, useful for pricing tables, specs, contact details, and product information.
• Research: multi-source research that breaks a question into sub-queries, reads sources, and returns a cited report.

How it works in NanoGPT

Sofya as a web search provider. Enable web search and select Sofya. NanoGPT calls Sofya, gets back page content, and injects it into the model's context. This is useful for recent information, fact-checking, and anything where a short snippet is not enough.

Sofya Research as a model. Select sofya-research from the model list or API. It runs Sofya's research endpoint and returns a structured report.

sofya-research

For API use with web search on any model:

gpt-4o:online/sofya

Benchmarks

Sofya publishes SimpleQA results on its homepage. In their June 2026 run:

• Agentic SimpleQA: 97% accuracy, #2 of 8 backends, 1.12 searches per question on average.
• Single-shot retrieval: 87.3%, #4, behind Parallel, Exa, and Claude's built-in WebSearch.
• Latency: snippets at 0.88s p50, basic search at 3.86s p50.

The distinction matters. Their agentic test used a search-and-reformulate loop; single-shot allowed one query only. Claude's built-in WebSearch also is not a standalone API, so it is more of a reference point than a direct replacement for Sofya in most stacks.

Pricing

Sofya charges credits at $0.005 each, with a free tier for eligible GitHub accounts.

NanoGPT prices include our standard 5% pass-through markup on Sofya's credit cost.

When to use it

Use Sofya when the answer depends on current web content and a snippet is not enough context: comparing products, tracking what changed in a release, pulling details from docs, or getting a cited report on a fast-moving topic. For quick one-off lookups, a lighter provider may be enough.

Sofya also publishes a native MCP endpoint, an API docs page, and an AI-readable skill file, which makes it directly usable in agent stacks like Codex, Cursor, Claude Code, Windsurf, and VS Code.

Try it

In the app: enable web search and select Sofya as the provider.

For research reports: select Sofya Research from the model list.