Library › Book 1 › Token Budget Analysis with Real Numbers
Token Budget Analysis with Real Numbers
10.1🔒Context Windows in July 2026Frontier models -- Claude Opus 4.8, OpenAI GPT-5.5, Google Gemini 3.1 Pro -- ship context windows from the 200K-token class up to 1M+. The…
10.2🔒The Headline Number FirstMicrosoft's published benchmark for a typical browser automation task:
10.3🔒MCP Token Breakdown: A Real Playwright SessionAn MCP server exposes 15-25 browser tools, each described by a JSON schema of roughly 150-280 tokens: roughly 3,000-4,000 tokens of schemas…
10.4🔒The Same Test via the Playwright CLISkill loading costs ~1,000 tokens once (SKILL.md) plus ~50 tokens per turn for the skill description.
10.5🔒The Honest CaveatPer-step arithmetic like the above can suggest ratios of 20-30x or more, and early CLI-skill advocacy -- earlier editions of this book…
10.6🔒Cost at ScalePer task, MCP consumes ~57% of a 200K window versus ~13% for the CLI; at an illustrative $15/M input tokens that is ~$1.71 versus ~$0.41…
10.7🔒Headroom, and When Token Cost Does Not MatterThe ~87K tokens saved per task are working room: source files, test definitions, error analysis, and history that would not otherwise fit…
10.8🔒Key Takeaways- The anchor numbers: ~114K tokens per typical task via MCP, ~27K via the Playwright CLI -- a measured ~4x, up to ~10x on long sessions…
10.9🔒Q&ASelf-Assessment Quiz1. What are the two Microsoft benchmark numbers to memorize, and what do they measure? 2. What is the "schema tax," roughly how large per…
10.10🔒Exercises[Beginner] Exercise 10.1: Using the per-component figures in this chapter, build the token ledger for a 10-step smoke test (one page, one…
10.11🔒Career Translation- Produced a token budget analysis for AI-driven browser testing anchored to Microsoft's measured benchmark (~27K vs ~114K tokens/task…
10.12🔒Q&AInterview Depth CheckPrompt: Walk through where the tokens go in a 20-step login test via MCP, and reconcile your estimate with Microsoft's ~114K measured…