The cheapest AI-generated Playwright test can become the most expensive test in the suite if an AI agent has to reason through it every time it runs.
That is the uncomfortable part of AI-assisted Playwright testing. Enterprise teams often want the speed of AI agents and Playwright MCP, while the organization quite reasonably ask two questions: how much will this cost, and can we trust what it produces?
Published data gives us enough evidence to examine both concerns. The Playwright MCP allows AI agents to interact with web pages through structured accessibility snapshots, which means every browser action can create more context for the model to process. The difference between a one-time authoring expense and a recurring execution bill becomes clearer once we look at where MCP tokens go and how the cost grows across a Playwright test suite.
The token bill starts before the test does
There’s no question whether Playwright’s AI capabilities are valuable. Playwright makes it easy to build entire test suites at scale in seconds. But like any open-source project, the ownership of governing the outcome of those tests lies with the user, and the decision of how to leverage Playwright becomes a more pressing question. The outcome includes not only speed to creation, but scalability, reliability, observability, and ultimately your organization’s ROI.
Pramod Dutta, principal SDET at Tekion Corp, an AI and LLM Test Architect, and the founder of TheTestingAcademy, published an independent comparison of Playwright MCP and the Playwright CLI using an eight-step login and dashboard task.
Dutta’s flow consisted of opening a staging app, logging in, navigating to the analytics dashboard, verifying three KPI cards rendered with non-zero values, clicking into a detail report, and screenshotting the resultt. Completing this flow with Playwright MCP consumed approximately 89,000 tokens. The same flow consumed roughly 24,000 through the Playwright CLI.
Dutta’s article also cites figures from a Playwright CLI launch comparison attributed to Microsoft, which measured one test consuming approximately 114,000 tokens through MCP against 27,000 for the same taskdone through the CLI.
The results suggest a meaningful cost pattern: MCP-driven browser work can use roughly four times as many context tokens as a leaner CLI workflow. Dutta attributes approximately 4,200 tokens to tool schemas loaded at the start of a session, around 3,800 tokens to a login-form snapshot, and approximately 12,000 tokens to a dashboard snapshot.
Larger enterprise pages can return much more, and a public issue opened by GitHub user jjcampana in Microsoft’s Playwright MCP repository reported a sixfold increase in token usage between versions 0.0.30 and 0.0.32 while solving the same task. That report reinforces the need to measure the exact MCP version, configuration, and application pages used by a team.
The economics change with repetition
During authoring, an agent may inspect the application, choose selectors, enter data, verify results, recover from a failed step, and revise the test. That process can consume tens of thousands of tokens, although it produces an asset that can then be reviewed and committed as ordinary Playwright code.
An MCP-driven execution loop has a different profile. Every run may require the agent to observe the page, interpret the returned structure, decide what to do next, and evaluate the result. A five-minute regression check can therefore become a recurring inference workload.
Now, multiply that across hundreds or thousands of tests, scheduled runs, pull requests, browsers, environments, and retries, and the recurring token volume can grow faster than the test count suggests.
From agent session to test asset
Teams that want agent-assisted authoring can reduce the recurring burden by separating how tests are created from how they run. Leapwork Play connects AI agents to Playwright-based validation through MCP while keeping execution governed and deterministic. The same input produces the same result on every run, with no model call at runtime.
Play adds role-based access, approvals, audit trails, and evidence-linked steps that trace results back to source requirements. Teams can also import existing Playwright tests and Git repositories, preserving their investment and avoiding a rewrite.
Reusable components and self-healing help reduce maintenance as applications change, with Leapwork reporting 50 to 70 percent lower maintenance costs. CI/CD integrations for Azure DevOps, GitHub Actions, and Jenkins keep validation connected to delivery, while visual debugging helps engineers understand what happened when a test fails.
Play’s Deterministic by Design approach of separating AI authoring and deterministic execution also strikes a key balance between coverage and validation. AI-assisted authoring can accelerate coverage, while deterministic execution makes recurring validation more predictable, easier to audit, and easier to budget.
