TinyFish Web Agent – 82% on Mind2Web vs Operator 43%

TL;DR

TL;DR: TinyFish is an enterprise browser automation platform that runs hundreds of web sessions in parallel using a serverless architecture, achieving 82% on the Mind2Web hard task benchmark versus OpenAI Operator 43%.

Source and Accuracy Notes

⚠️ This section is MANDATORY. All links must be verified from actual source, not guessed.

• Project page: tinyfish.ai — verified 2026-08-24
• Blog / benchmark: tinyfish.ai/blog/mind2web — verified 2026-08-24
• GitHub: github.com/tinyfish-io — verified 2026-08-24
• License: not confirmed — no open-source license found on GitHub org; this is a commercial SaaS product
• HN launch thread: news.ycombinator.com/item?id=46991520 — 17 points, 2026-02-12

What Is TinyFish?

TinyFish is a browser automation platform built specifically for AI agents. Rather than running a single browser instance on your laptop, it orchestrates hundreds of browser sessions concurrently on cloud infrastructure. The product is aimed at enterprises that need to deploy AI agents capable of navigating, authenticating, and extracting data from multiple websites simultaneously.

The company was founded by Sudheesh Nair, Shuhao Zhang, and Keith Zhai, and is based in Palo Alto.

Architecture: Serverless Browser Orchestration

The core problem TinyFish solves is the infrastructure overhead of browser automation at scale. Traditional browser agents run on a single machine — a self-hosted agent might launch a headless Chrome instance and control it via Playwright or Puppeteer. This works for single-task workflows but breaks down when you need to:

• Run the same task across hundreds of domains in parallel
• Maintain authenticated sessions across multiple websites
• Scale from 10 to 10,000 concurrent browser sessions without managing a fleet of VMs

TinyFish’s architecture is described as fully serverless. Browser sessions run in isolated cloud environments and are orchestrated via an API. The platform handles session lifecycle, proxy rotation, and state management without you managing underlying compute.

Benchmark Performance: Mind2Web Hard

TinyFish published benchmark results on Mind2Web, a widely-used evaluation for web agents. Mind2Web tests AI agents on real-world web tasks drawn from actual websites — no synthetic or simulated environments.

Reported results (from TinyFish blog):

| Task Difficulty | TinyFish | OpenAI Operator | |---|---|---| | Easy | — | — | | Medium | — | — | | Hard | 82% | 43% |

The hard task category covers multi-step workflows requiring sustained context across pages, form-filling with validation, and handling dynamic content.

Important caveats:

• These numbers are self-reported by TinyFish on their own blog post. The methodology and exact evaluation setup are not independently verified. Other web agent frameworks have also published Mind2Web scores — for context, the OpenAI Operator evaluation report covers different benchmarks.
• The TinyFish benchmark post does not name a specific TinyFish model or configuration, so it’s unclear whether these results use a proprietary model, an API provider, or an open-source agent framework as the underlying LLM.

Verify the benchmark methodology directly at tinyfish.ai/blog/mind2web before citing these numbers in external contexts.

Key Features

From the product page and GitHub org:

• Concurrent session management — orchestrate hundreds of browser instances simultaneously via API
• Serverless execution — no VM fleet to manage; sessions scale elastically
• Authentication persistence — maintain logged-in state across sessions without re-authenticating per task
• Proxy rotation — built-in support for residential and data center proxies
• Data extraction — structured output from web pages (HTML, JSON, screenshots)
• JavaScript rendering — handles SPAs, client-side rendered content, and dynamic DOM manipulation
• GitHub presence — github.com/tinyfish-io — check individual repos for open-source components

Use Cases

Typical deployments include:

• Price intelligence — monitor competitor pricing across hundreds of e-commerce sites in parallel
• Lead enrichment — automatically visit and extract contact data from company directories
• Market research — aggregate product listings, reviews, and ratings from multiple platforms
• Form automation — submit data to web forms across multiple targets simultaneously
• Authenticated data extraction — extract data from pages behind login walls

Pricing

No public pricing is available on the website. The product appears to be enterprise-focused with custom pricing based on session volume. Open a sales inquiry at tinyfish.ai for current rates.

How It Compares

| Feature | TinyFish | Skyvern | Browser Use | OpenAI Operator | |---|---|---|---|---| | Serverless architecture | Yes | No (self-hosted) | No (self-hosted) | API only | | Concurrent sessions | Hundreds | Single | Single | Single | | Mind2Web hard score | 82% | — | — | 43% | | Open source | Unclear | Yes (Apache 2.0) | Yes (MIT) | No | | Enterprise pricing | Yes | Self-hosted | Self-hosted | Usage-based |

Practical Evaluation Checklist

If you’re evaluating TinyFish for your workflow:

• [ ] Identify the underlying LLM — the blog does not specify which model powers the agent
• [ ] Clarify what “serverless” means in their infrastructure (AWS Lambda? Cloudflare Workers? custom runtime?)
• [ ] Request a trial or sandbox to test against your specific target websites
• [ ] Compare per-session cost vs running equivalent Playwright/Puppeteer infrastructure
• [ ] Check if any open-source components are published on github.com/tinyfish-io
• [ ] Evaluate whether you need concurrent session orchestration or a single-agent approach suffices

Security Notes

• Browser sessions run in isolated cloud environments — confirm with TinyFish what the isolation boundary is (container, VM, sandbox)
• Authentication credentials are stored server-side for session persistence — evaluate the credential storage and rotation policy
• Review their data retention and session log policies, especially if automating behind login walls
• For compliance-sensitive workflows, request a security questionnaire and SOC 2 or equivalent report

FAQ

Q: Is TinyFish open source? A: The GitHub org (github.com/tinyfish-io) exists but no public repositories with open-source licenses were found as of 2026-08-24. TinyFish appears to be a commercial SaaS product. Check their GitHub for any future open-source releases.

Q: What LLM powers the TinyFish agent? A: The benchmark post does not specify. Contact TinyFish directly for the model configuration and API details.

Q: How does TinyFish handle CAPTCHAs and anti-bot measures? A: The product page mentions proxy rotation as a feature. Specific anti-detection capabilities (Stealth mode, fingerprint randomization) are not documented on the public site.

Q: Can I self-host TinyFish? A: The architecture is described as fully serverless and managed. There is no documented self-hosted deployment option as of this writing.

Conclusion

TinyFish targets a specific gap in the browser automation market: enterprise-scale, concurrent web agent infrastructure. The Mind2Web hard benchmark result (82% vs Operator 43%) is a notable data point, though it comes directly from TinyFish’s own blog and should be verified against independent evaluations before citing it in adversarial contexts.

If your workflow requires a single AI agent to browse the web, open-source options like Browser Use or Skyvern running on your own infrastructure offer more transparency and cost control. If you need to orchestrate hundreds of concurrent web tasks with managed infrastructure, TinyFish is worth evaluating — but clarify the underlying model, pricing, and data handling before committing.

Source: tinyfish.ai | Benchmark source: tinyfish.ai/blog/mind2web | HN: news.ycombinator.com/item?id=46991520