Opus vs Sonnet: what's the real gap in your Claude Code bill?

It's the real rate card, not a rounding. Under Sonnet 5's current introductory pricing (through 2026-08-31), Sonnet lands at exactly 40% of Opus 4.8 on every axis - input, 5-minute cache write, cache read, and output - so a token costs 60% less the moment it runs on Sonnet instead of Opus. See the full input/output/cache-write/cache-read rate card in the Claude Code cost-per-token breakdown.

Because 39.8% is a share of turns, not a share of cost, and this project's own model-routing research found those Sonnet-safe turns carry only about 14.4% of a session's output tokens - easy turns are cheap turns. Worse, modeling per-turn routing honestly (with the real cache-miss penalty a model switch triggers) actually loses money: -122%. The naive number this tool shows next to the honest one is there specifically to make that gap visible, not to sell you on it.

Yes, but only through cache-safe coarse routing - a whole subagent pinned to Sonnet or Haiku with its own separate cache, not a switch mid-thread. That's the same mechanism the subagent cost calculator prices. The realistic envelope for that lever, measured across this project's own research, runs low single digits up to roughly 10-15% of total spend, materially larger only if you're currently defaulting everything to Opus.

100% is a reasonable default, not a scare tactic - Claude Code custom subagents default to Opus unless a frontmatter model field or CLAUDE_CODE_SUBAGENT_MODEL override says otherwise, and most setups never add one. If you already run some work on Sonnet or Haiku, lower the slider; the honest-savings math scales down with it, since there's less Opus spend left to shave.

No - it's a modeled comparison built from Anthropic's published per-token rates and this project's own measured session composition (1,037 real sessions), not a reading of your actual transcripts. Your real split depends on which subagents and models you actually invoke. Run the free UsageCut scan for your own numbers instead of an estimate.

Yes. The scan and every fix are free - no card, no subscription, no signup. You only give an email if you want the optimization plan and one-command undo sent to your inbox.

The scan runs entirely on your machine. It never uploads your code, never reads your API key, and never sends your prompts or conversations anywhere. Only anonymous summary counts leave, and only if you choose to share them.

It reads your local setup and session history to find waste - idle MCP servers loaded into every session, an oversized CLAUDE.md re-sent on every request, duplicate file reads, and bloated tool output - then trims it losslessly, so each session carries less context and you hit your limit far less often.

ClockedCode is the curated, conflict-free Claude Code setup from the same maker. UsageCut makes your current setup leaner; ClockedCode makes it genuinely better - a set of hand-picked tools, specialist agents, and a tuned CLAUDE.md, installed in one paste.

Developers who use Claude Code daily and want a setup that is powerful out of the box instead of spending hours researching and wiring up plugins, agents, and MCP servers themselves.

A vetted, conflict-free bundle of hand-picked tools, specialist agents, and a tuned CLAUDE.md - everything installed together in one paste, with no setup archaeology.

Yes. ClockedCode is a one-time purchase with an instant download - no subscription and no recurring fees.

Yes. It installs alongside what you already have and is fully reversible. Run UsageCut first to clean your current setup, then add ClockedCode to level it up.