AI Coding Costs for Legacy vs Greenfield Projects: A Real-World Budget Guide

Why Legacy Code Is More Expensive to AI-Assist

When developers estimate their AI coding budget, they typically think in terms of model price per token. What they underestimate is how dramatically the type of codebase affects the total token bill. Legacy projects — codebases that are 5-15+ years old, have accumulated technical debt, use older frameworks, and lack comprehensive documentation — drive AI costs in ways that greenfield projects simply do not.

The core mechanisms are straightforward: legacy code requires more context to understand, has worse prompt caching efficiency because the context changes frequently, and typically requires more debugging iterations before generating correct output. Each of these factors multiplies your token spend relative to a clean greenfield project on identical functionality.

Context Window Requirements: Legacy vs Greenfield

In a greenfield project, the relevant context for any given task is usually small and well-organized. You are adding a new feature to a codebase with consistent patterns, clear naming conventions, and interfaces you designed yourself. A typical AI coding prompt might include 5-10K tokens of context.

In a legacy project, that same feature addition might require the model to understand 20-60K tokens of context: the existing module's tangled dependencies, the undocumented global state it relies on, the three different patterns previous developers used before settling on the current approach, and the comments explaining why a seemingly obvious solution was tried and abandoned. This context inflation multiplies directly into higher input token costs.

The Prompt Caching Problem in Legacy Codebases

Prompt caching works by reusing the processed representation of input tokens that appear at the beginning of your context. If you send the same large system prompt or codebase context repeatedly, caching can reduce your effective input cost by 80-90%. This is extremely valuable in greenfield projects where your project structure and documentation are stable.

Legacy projects break caching efficiency in two ways. First, the context itself changes more frequently — you are working across different modules with different dependencies in each session, preventing the system prompt from staying cache-warm. Second, legacy code often requires injecting specific file contents mid-context rather than at the top, which places uncacheable tokens before the potentially cacheable sections.

A greenfield project with 70% cache hit rate on a 30K token context effectively costs you 9K tokens per request in input. The same legacy project with 25% cache hit rate on a 50K context costs 37.5K tokens per request — a 4x difference in effective input cost from caching alone.

Budgeting by Project Type

Using Claude Sonnet 4.6 at $3.00/$15.00 per million tokens as a reference:

Strategies to Cut Legacy Codebase AI Costs

You cannot change the codebase instantly, but you can reduce the token overhead of working with it:

• Create a lightweight documentation layer: A concise module index (1-2K tokens) that describes what each file does and its key interfaces reduces the amount of raw source code you need to inject as context per task.
• Isolate the task scope aggressively: Resist the temptation to include entire files when only a function matters. Precise context reduces tokens and often improves model output quality by reducing noise.
• Use a budget model for exploration, frontier for implementation: Use DeepSeek V4 Flash ($0.14/$0.28) to understand the legacy code and identify the right approach, then use Sonnet or Opus only for the actual implementation step.
• Write tests first: For legacy code, specifying the expected behavior via tests reduces ambiguous back-and-forth. The model generates code against a clear target rather than inferring intent from tangled existing code.

Use the AI Cost Estimator to set a monthly budget based on your project type and expected feature count. Legacy codebase work warrants a 3-5x buffer over what you would budget for equivalent greenfield work.