The same CLAUDE.md does not cost the same on Opus, Sonnet, and Haiku.

Every other guide on this topic looks at drift over time, drift across config files, or drift between teammates. This one is about drift across the thing most engineers change three times a day with a slash command and never think about: the model itself.

You wrote one CLAUDE.md. You read one cost number. You switch from Opus to Sonnet for routine edits, back to Opus for the planning step, drop to Haiku for the boilerplate. The file does not move. The bill does, by 5x at every step. So does the rate at which the rules in the file are actually obeyed.

1. The one line in the analyzer that locks the cost to Opus

The free analyzer at ccmd.dev is a pure function. You paste a CLAUDE.md, it returns a snapshot. Two of the numbers in that snapshot are model-independent (totalTokens, the Karpathy rubric pass rate). The dollar number is not. Here is the math the analyzer runs, verbatim:

Read line 267. The constant OPUS_IN_PER_M = 15 is the entire cross-model assumption baked into the tool. It reflects an older Opus input rate. The current 2026 Opus 4.7 input rate is closer to $5 per million; Sonnet 4.6 is $3; Haiku 4.5 is $1. If your team standardized on Sonnet a month ago and you are still quoting the analyzer's default cost, you are quoting a number that is roughly five times your actual input bill.

We will land a model dropdown on the homepage analyzer in the next release. Until then, the number you see is the Opus row of a larger table.

2. The spread on a single file, four ways

Same CLAUDE.md, 6,420 tokens. Same session length, 30 turns. Four input prices. The right column is the per-session input floor for the file alone, before any code reads, tool calls, or generations.

The bottom row is the most interesting one for vibe-coded budgets. A 6,420-token CLAUDE.md you would never write more than one of costs 19 cents per 30-turn session on Haiku for the rehydration alone. The same file on the analyzer's default Opus row reads as $2.89, fifteen times the actual Haiku bill. The fear that drives most CLAUDE.md compaction efforts is the Opus number. The number on your invoice is one of the other three.

3. The rubric score is model-independent. Observed adherence is not.

The Karpathy 12-rule rubric (see /karpathy-rules) scores whether your CLAUDE.md contains each rule. It does not score whether the model actually follows the rule across sessions. Those are different numbers. The first depends on what you wrote. The second depends on which model is reading it.

Short, surgical rules (R1 Think Before Coding, R7 Tests as Truth) hold across all three Claude models we tested. Long aspirational rules erode as the model shrinks. Here is the observed adherence on the same CLAUDE.md across 50 sessions per model:

The file scores 7/12 on the rubric (a property of the file). On Haiku, observed adherence drops to 3/12 (a property of the interaction). The four rules that drop out (R3, R5, R6, R11) all share a shape: long, abstract, no concrete failure mode named. Those are the lines that need rewriting if you plan to use Haiku for any non-boilerplate work, not deleting.

4. Opus to Haiku, dimension by dimension

The thing that drifts is not just dollars. Cost, adherence, cache behavior, and adaptive-thinking interaction all change in the same model switch. Six of them, side by side:

None of these rows is a flaw in any model. They are the cost of having three models to choose from and one CLAUDE.md to feed all three. The fix is not to write a different CLAUDE.md per model; that defeats the point. The fix is to know which row you are on when you look at any cost or adherence number.

5. The 30-line workaround until the dropdown ships

Drop this into your repo at scripts/claudemd-cost.ts. It reads your CLAUDE.md, takes a model name on the command line, and prints the right per-session input floor. It is the entire per-model cost adjustment for the free analyzer, written by hand so you do not wait for the next release:

One command, one number, no normalization assumption. If your team standardized on Sonnet, the third row of the price table is the only one that matters and the script prints exactly that. If you switch models four times a day, run the script four times and decide which baseline is honest to quote.

6. The honest rewrite: scope rules by intended model

If your team genuinely uses three models, the right move is not one CLAUDE.md with everything in it. It is a base CLAUDE.md plus two scoped sub-files, with a one-line gate at the top of each rule that says which model it is for. Anthropic's nested CLAUDE.md mechanic is the existing tool for this; see the nested CLAUDE.md token scope guide for the file-layout half and the multi-file token budget guide for the per-PR budget half. The short version: keep the rules that hold on Haiku in the root CLAUDE.md, push the long-form planning prose into a Skill that loads only when Opus picks it up, and stop pretending one file fits three models.

That is also the only configuration where the analyzer's default Opus number is honest. If the long rules only fire when Opus is the model, the Opus rate is the right rate to multiply by. The drift stops being a hidden 5x; it becomes a deliberate design.