Optimize claude-sonnet cost with Claude Code

Sonnet is the default model. Most Claude Code users are on it whether they know it or not. It is not expensive per token, but session depth compounds fast.

What claude-sonnet costs per session

Claude Sonnet pricing (as of April 2026) is $3.00 per million input tokens. A typical Claude Code session runs 30 turns. Each turn re-sends the full conversation context: tool results, file reads, prior messages, everything that came before it. By turn 30, the average context per turn reaches 50,000 tokens.

Why claude-sonnet bills compound

The compounding happens because of how the Claude API works, not how much Sonnet costs per token. Every request re-sends the entire conversation from the beginning. Turn 1 sends 1 turn of context. Turn 10 sends 10 turns. Turn 30 sends 30 turns. The total input tokens for a 30-turn session is not 30 × turn-1 tokens, but the sum of all growing contexts across every turn.

This is not a Sonnet-specific behavior. It applies to every Claude model. But because Sonnet is the default, and because developers run it for hours without thinking about session boundaries, the cost accumulates invisibly. The bill arrives at the end of the month, not after each session.

How The Distillery reduces claude-sonnet spend

The Distillery intercepts each request before it reaches Anthropic. It distils repeated context: deduplicated tool results, redundant file reads, verbose shell output, and forwards a smaller payload. Your session continues normally. Claude still gets everything it needs. You pay for less of what it already had.

Real-world cost reduction is 30-60% depending on session patterns. The conservative deterministic benchmark floor on Sonnet sessions is 20% on standard usage — heavy agentic sessions (long refactors, multi-file edits, repeated test runs) produce higher reductions because there is more repeated content to distil.

See the exact methodology: benchmark results →