Can AI solve a CAPTCHA? Partly.

In 2021–2024 studies, specialised models matched or beat people on image CAPTCHAs, hCaptcha included. AI agents lag: the best solved 40% in 2025, and OpenAI (2025) and Anthropic say theirs are built to hand them to you.

What the research measured

Below are the numbers and the vendors’ own words, with dates, then how an agent you run can get an hCaptcha token. Four published studies cover the question from both sides: solvers built for a single CAPTCHA, people solving CAPTCHAs that sites actually deploy, and general-purpose agents dropped in front of puzzles they were not trained on.

The first three rows are specialised solvers. Hossen and Hei (2021) built one for hCaptcha’s image challenges and ran it from a small container with no GPU. Searles et al. (2023) measured people on deployed CAPTCHAs and compared them with attacks in the literature: “bots can outperform humans, both in terms of solving time and accuracy, across all these captcha types.” The ETH Zurich reCAPTCHA v2 study (2024) solved every challenge and also found “no significant difference in the number of challenges humans and bots must solve to pass”.

The last row is the one most people mean today when they ask about AI. Open CaptchaWorld (May 2025, NeurIPS 2025) tested multimodal agents in a real browser on interactive, multi-step puzzles. People scored 93.3%; the best agent, Browser-Use driving OpenAI’s o3, scored 40.0%. The authors note that these models do well on static perception tasks but struggle when a puzzle needs several steps of looking, reasoning and clicking.

Can ChatGPT or Claude solve a CAPTCHA?

Both companies say their browser agents are built to hand CAPTCHAs back to the person. OpenAI’s Operator announcement (23 January 2025) says Operator “is trained to proactively ask the user to take over for tasks that require login, payment details, or when solving CAPTCHAs”. A note added on 17 July 2025 says Operator is now part of ChatGPT as ChatGPT agent.

Anthropic’s Claude in Chrome documentation for Claude Code says: “When Claude encounters a login page or CAPTCHA, it pauses and asks you to handle it manually.” Open CaptchaWorld ran the underlying models, among them GPT-4o, o3, Claude 3.7 Sonnet and Gemini 2.5 Pro, through the Browser-Use agent instead, with no such hand-off, and 40% was the best any of them reached.

What hCaptcha says about AI

hCaptcha has written about AI solvers for years, and its argument has stayed the same: the answer is only part of the test.

• How, not only whether. In “Why CAPTCHAs Will Be With Us Always” (May 2021) it says it analyses “how the challenge was answered, not just whether it was correct”, and frequently changes both the questions and the data.
• Different mistakes. “Detecting Large Language Models” (April 2023) says “AI/ML systems make different kinds of mistakes than people”, and that hCaptcha uses this to detect LLMs and tell which model gave an answer.
• Moving target. A post from 23 February 2026 says that if a challenge “starts to see higher automation solve rates, it is rapidly discarded or updated”, that vision-language models still show gaps against people “with simple tasks that take people only a few seconds”, and that “a browser agent is only as good as its model”.
• Agents should say what they are. The same post says legitimate agents “should clearly identify themselves”. In October 2026 hCaptcha Enterprise added Web Bot Auth for signed agent traffic and a WebMCP verification tool that returns requires_human when an interactive challenge opens, at which point “the agent must stop and ask the person who initiated the task to complete it.”

Many visitors never see a puzzle at all. The February 2026 post says hCaptcha “has always supported passive and invisible abuse detection” and that most visitors checked by its Pro and Enterprise plans “experience zero friction”; why am I getting hCaptcha covers what makes a visit get a challenge.

A right answer is not a token

What a site needs from a CAPTCHA is a token: for hCaptcha, a P1_ string in the form’s h-captcha-response field, which the site’s server checks with hCaptcha’s siteverify before it accepts the form. A model that picks the right tiles has done one step. hCaptcha still decides whether to issue a token, and the site still decides whether to accept it. That is why the research numbers above, which count correct answers, do not translate directly into an agent getting through a form. The hCaptcha token guide covers the format and its two-minute lifetime.

Where NoneCap fits

If you run an agent on sites where you are allowed to automate, the practical answer is to stop asking the model to solve the puzzle. Give it a tool instead. NoneCap is an hCaptcha solving API: the agent sends the sitekey and page url, gets back a real P1_ token, puts it in h-captcha-response and carries on. The agent spends no steps on tiles, and failed, cancelled and expired solves are never charged.

In Claude Code, Cursor or another MCP client, the NoneCap MCP server adds the tool with one line:

claude mcp add nonecap --env NONECAP_API_KEY=<your key> -- npx -y -p nonecap@latest nonecap-mcp

For a custom agent loop, the AI agents guide shows the tool definition, a handler that calls POST /v1/solves, and the inject step. NoneCap solves hCaptcha only; it does not help with reCAPTCHA, Turnstile or other CAPTCHAs.

Set up hCaptcha for your agent →