Claude Code has real costs. For most QA engineers, those costs are modest relative to the time saved — but "modest" depends heavily on how you use it. This lesson covers the billing model, the practices that keep costs predictable, the workflows worth avoiding, and a realistic picture of where Claude Code fits (and does not fit) in day-to-day QA work.
How billing works
Subscription mode (Claude Pro or Team via claude.ai OAuth): flat monthly rate, Claude Code usage is included. This is the most common setup for individual QA engineers and small teams. Predictable cost, no per-session surprise.
API mode (API key, no subscription): pay-per-token. Scales with usage. Right for organisations that want metered billing, need to manage Claude Code access across many engineers, or run it in CI where subscription terms do not apply.
For most QA teams, a Team subscription is more predictable than API billing. Once you have a sense of your usage patterns from the first month, you can decide whether API mode would be cheaper.
Token cost reference (API mode, rough estimates)
| Task | Tokens | Approx. cost at Sonnet pricing |
|---|---|---|
| Single test generation | 5K–20K | $0.03–$0.15 |
| Bulk refactor of 10 files | 50K–150K | $0.40–$1.20 |
| Framework migration (50 tests) | 300K–800K | $2.50–$6.50 |
| Full suite audit + triage | 500K–1.5M | $4.00–$12 |
These are approximations that depend on file sizes and prompt verbosity. The first time you run a large task, use /cost before and after to calibrate.
Cost-saving practices that matter most
Default to Sonnet. Sonnet 4.6 handles the vast majority of QA tasks — test generation, refactoring, debugging, Page Object creation — at roughly half the cost of Opus. Switch to Opus with /model only when output quality from Sonnet has been genuinely insufficient for a specific complex task.
Scope prompts precisely. "Look at everything and find issues" causes Claude Code to read the entire codebase. "Read only tests/checkout/ and identify timing anti-patterns" reads a fraction of that. Precise scope = fewer tokens.
Use CLAUDE.md. Every session that starts without CLAUDE.md has to re-establish project context through back-and-forth. CLAUDE.md provides that context upfront in one read — consistently cheaper than ad-hoc explanation.
Continue sessions rather than restarting. claude --continue resumes the last session with its conversation context intact. Starting a fresh session on the same task discards context that was already paid for and has to re-read files from scratch.
End sessions when done. Long-running sessions accumulate context cost as the conversation grows. Close and reopen for a new task rather than letting a session run across unrelated work all day.
The cost-vs-value matrix
Not every QA task is worth routing through Claude Code. The filter is simple: high-volume or high-complexity tasks where Claude Code is significantly faster than manual work justify the cost. Low-complexity, low-volume tasks often don't.
Claude Code cost-vs-value decision framework
| Low cost | High cost | |
|---|---|---|
| High value | Always use — bulk refactors, test generation for new features, POM scaffolding, debug analysis | Use selectively — large migrations, full suite audits. Check ROI before starting. |
| Low value | Optional — small one-off fixes where manual is just as fast. Fine either way. | Avoid — trivial changes you understand completely. Faster to just type it. |
When not to use Claude Code
Trivial changes. Renaming a variable, fixing a typo, adjusting a single timeout value — these take longer to prompt than to type. Use your editor.
Things you do not understand. You cannot review what you cannot evaluate. If you lack the context to judge whether Claude Code's output is correct, get the context first.
Time-critical production hotfixes. When a production incident is live and every minute counts, reach for your established, proven manual process. An agentic tool that needs your review is not the right tool for a 90-second fix.
Sensitive environments with compliance requirements. Claude Code sends context to the Anthropic API. If your project involves regulated data — HIPAA, PCI, financial records — verify your compliance posture before using Claude Code on that codebase.
Realistic productivity trajectory
The first month of using Claude Code is slower than you expect. You are learning how to write effective prompts, building your CLAUDE.md, and developing an instinct for what Claude Code handles well. Expect roughly the same productivity as before, with occasional faster wins.
By month three, the patterns are established. Test generation, bulk refactoring, and debugging analysis are meaningfully faster. Engineers who tracked time report 2–3x speedups on the specific tasks where Claude Code fits.
By month six, it is integrated into the workflow as naturally as the IDE and git. The productivity gains compound because CLAUDE.md keeps improving, the prompt library grows, and the team shares patterns.
⚠️ Common Mistakes
- Defaulting to Opus for everything. The cost difference between Sonnet and Opus is significant at scale. Use Sonnet as the default and keep a mental note of specific task types where Sonnet's output quality was insufficient — those are the exceptions that warrant Opus.
- No cost monitoring in the first month. The first 30 days establish your usage baseline. Check the Anthropic Console weekly so you know what "normal" looks like before costs compound.
- Using Claude Code for tasks where you cannot evaluate the output. The ROI of AI assistance depends on your ability to review the result. If you cannot judge whether a generated test is correct, you are not saving time — you are deferring risk.
🎯 Practice Task
Audit your Claude Code usage patterns. 15 minutes.
- Think back over the last week of using Claude Code. List three tasks you used it for.
- For each one, apply the cost-vs-value matrix: was it high-value? Was the cost justified?
- Identify one task type you have been routing through Claude Code that would be faster manually.
- Identify one task type you have been doing manually that Claude Code would handle significantly faster.
- Adjust your default workflow accordingly.
Chapter 6 puts everything from this course into practice — building a complete Playwright test suite for a real application, using Claude Code as your primary tool from project setup through documentation.