Kiwi Bench
Independent benchmarks of the models, data providers, and agent tools your agents call.
Test any model, data provider, or agent tool on your use case. Keep the results. Call the winner with one key.
Document parsing
Text, tables, and key fields from invoices, contracts, filings, scans, forms, and slides.
Reducto, Unstructured, Mistral OCR, Azure Document Intelligence, Google Document AI, AWS Textract, Docling, Marker
Models
Extraction, classification, summarization, and tool calls, across 16 chat and reasoning models.
Anthropic, OpenAI, Google, DeepSeek, Qwen, Moonshot, Z.ai, Mistral, Meta
Company enrichment
Headcount, HQ, industry, funding, tech stack, and open roles for 200 companies, checked by hand.
People Data Labs, Crustdata, Coresignal, TheirStack
Cloud browsers
Whether a page loads and the content is there, on news, ecommerce, docs, government, and bot-protected sites.
Browserbase, Steel, Kernel, Hyperbrowser, Browserless, Anchor Browser
How they are run
- Same inputs, same client, same region. Every provider gets the identical task set, called from the same harness in the same cloud region. The region is recorded with each run.
- Documented defaults. Each provider runs on its documented default or recommended settings. Any setting we change is listed on the results page.
- Providers see their results first. Five business days before a page goes up, each provider gets its own results and the method. Genuine misconfigurations are fixed and noted in the changelog.
- Dated and priced. Every page shows the run date, the API versions, the prices used and where they came from, and the known limitations.
Kiwi offers access to these providers through one API key. No provider pays for its ranking.