TinyFish launched a unified web infrastructure platform that consolidates search, browser automation, and content extraction under a single API key. The company claims their system achieved 90% accuracy on the Mind2Web benchmark, outperforming Gemini by 21 points, OpenAI by 29, and Anthropic by 34 across 300 tasks. Their platform handles multi-step workflows, form interactions, and JavaScript-heavy sites while managing proxies and stealth profiles automatically.
This tackles a real pain point I've seen repeatedly—teams juggling Puppeteer for browser automation, separate search APIs, content extraction services, and proxy management just to get agents working with real websites. The fragmentation is brutal, especially when you need agents to navigate complex user flows or extract data from sites that weren't built for programmatic access. TinyFish positioning themselves as "the same infrastructure used by Google, DoorDash, and ClassPass" suggests they've been operating under the radar serving enterprise clients.
What's interesting is their cookbook approach—they're publishing open-source examples and running a $2M accelerator program, clearly betting on developer adoption over just enterprise sales. The Mind2Web benchmark claims are bold but specific enough to verify. However, web automation promises always sound better in demos than production. Real websites break, change layouts, add CAPTCHAs, and throw edge cases that make even the best automation brittle.
For developers building AI agents, this could eliminate significant infrastructure overhead—if it actually works reliably. The natural language goal approach ("URL + plain English, get structured JSON back") is exactly what agent builders need. But given how many companies have promised unified web automation platforms, I'd want to see sustained performance across diverse sites before betting production workflows on it.