What does your CLAUDE.md cost on free models?
Search this question and you get two kinds of pages. One kind explains the free tiers (how many messages you get, how big the free credit is) but never mentions your config file. The other kind explains what a CLAUDE.md costs, then quietly assumes you are paying a per-token dollar rate and goes silent the moment the model is free. Nobody connects the two. So here is the connection.
The cost is a token count, and tokens do not know the price
The cleanest way to see why a free model does not make your config free is to read the cost code in ccmd's own analyzer. The function takes a single string. There is no argument for "and what model are you on."
Look at the two lines that matter. The cost itself is estimatedTokensFireEveryTurn = totalTokens (line 264), a pure token count with no price in it. The dollar figure underneath only multiplies that count by a rate (OPUS_IN_PER_M = 15). Set the rate to 0 for a free model and the dollar figure collapses to zero. The token count above it does not move. That is the whole argument: the model sets the multiplier, the file sets the tokens, and the tokens are the part you actually carry on every turn.
The token count itself comes from one line, the same chars / 4 heuristic every CLI uses:
A 24,000-character CLAUDE.md is about 6,000 tokens. That number is the same on Opus, on the free Sonnet 4.5 default, on a Cursor free request, and on a local model you run for nothing. The model never enters the calculation.
Free does not mean costless. It means a different unit.
On a paid model the cost shows up as dollars. On the free tiers people actually use, the same token count shows up as one of three rationed budgets. Here is how the identical 6,000 tokens land in each:
| Feature | How CLAUDE.md tokens hit it | Free budget you spend |
|---|---|---|
| Claude free plan (claude.ai, Sonnet 4.5 default) | Config text rides inside every message's context; the cap counts messages, the window holds 200K tokens. | ~15–40 messages / 5h window |
| New Anthropic API account (free starter credit) | Billed at the real per-token rate against the free balance. A bloated file empties the credit in fewer sessions. | A few Claude Code sessions of credit |
| Cursor Hobby (free) tier | Cursor now meters by token consumption, so the tokens in your .cursorrules / AGENTS.md count against the hidden monthly limit. | 50 requests + 2,000 completions / mo, token-metered |
| Paid model (Opus 4.7) | Same token count as every row above. The only thing that changes between free and paid is the multiplier. | $5 / M input tokens |
The bottom row is the paid model, dropped in on purpose. Read down the left column and the token count never changes. The thing labeled "free" is the multiplier, not the file. Whether that multiplier is dollars, messages, credit, or a hidden request meter, your config is spending it on every turn before you have asked a question.
What the analyzer reports when the model is free
Paste a 6,000-token file into ccmd.dev and the score reads the same on a free model as on Opus, because the headline number was never a dollar figure. The dollars are a footnote on the token count:
The savings line is the point. The analyzer flags the same lines as bloat, vague, or cache-busting whether you are paying dollars or spending a free quota, because those findings are properties of the text, not the bill. The cache_bust finding (a date stamp or "today is" near the top of the file) matters even on free tiers that support prompt caching, because a busted cache means full-price token processing on every session.
The one case where free really is free of dollars
A local model you run yourself (Ollama, llama.cpp) has a true marginal cost of zero dollars per token. No card, no credit, no message cap. That is the only honest "free model" where the dollar answer is a flat zero with no asterisk.
Even there, the config is not free. Every token of CLAUDE.md still has to be processed on every turn, which costs you latency and a slice of the context window. Local models tend to ship smaller context windows than the hosted frontier models, so a 6,000-token config eats a larger share of what you have to work with. The dollar cost is zero; the token cost is not, and the token cost is what slows the model down and crowds out the code you actually want it to read.
See what your config costs in whatever unit you pay in
Paste your CLAUDE.md, AGENTS.md, or .cursorrules at ccmd.dev for a free token score, or book 15 minutes and we will walk through which lines are draining your quota and which ones to cut first.
FAQ
Frequently asked questions
Does CLAUDE.md actually cost money on a free model?
No dollars, yes budget. A free model charges $0 per token, so the dollar line on any cost calculator reads zero. But CLAUDE.md sits near the top of the system prompt and gets re-sent on every turn, so it still consumes the scarce resource each free tier rations: your message cap, your free API credit balance, or your token-metered request quota. The token count is byte-for-byte identical to what a paid model would bill. Only the unit of measurement changes.
How does the ccmd analyzer prove the cost is the same on free and paid?
Open src/lib/analyzer.ts. The function is analyzeConfig(input: string) on line 134. It takes one string and has no second argument for which model you are on. The cost line on line 264 is estimatedTokensFireEveryTurn = totalTokens, a pure token count. The dollar figure on lines 266 to 269 only multiplies that count by a fixed rate (OPUS_IN_PER_M = 15). Set that rate to 0 for a free model and the dollar figure collapses to zero, while estimatedTokensFireEveryTurn does not move a single token. The token cost is model-independent by construction.
If I am on the claude.ai free plan, does my CLAUDE.md even load?
CLAUDE.md is a Claude Code file, not a chat file, so the free claude.ai web chat does not auto-load it. The cost there shows up only if you paste config text into a Project's custom instructions, in which case that text rides inside every message's 200K-token context window. The free plan caps you at roughly 15 to 40 messages per 5-hour window, so the cost is paid in messages and context space, not dollars. If you run Claude Code itself, you are on a Pro/Max subscription or the API, which is where the credit and quota math below applies.
What about a new Anthropic API account with the free starter credit?
New API accounts get a small free credit balance, enough to test an integration and run a few Claude Code sessions. That credit is spent at the real per-token rate, so a bloated CLAUDE.md drains it in fewer sessions than a tight one. This is the closest thing to a true free model cost for Claude Code: the file is billed exactly as it would be on a paid account, the difference is only that the bill comes out of a free pool instead of your card. Trim the file and the same pool lasts longer.
Cursor's free tier is request-based, so does file size matter there?
It used to be purely request-count, but Cursor moved to token-based metering, and the dashboard no longer shows how much of the hidden monthly limit you have used. That means the tokens in your .cursorrules or AGENTS.md count directly against the quota every time the agent reads them. A multi-step Agent task can already burn 5 to 10 requests; a heavy rules file on top of that pulls you toward the limit faster. ccmd scores .cursorrules and AGENTS.md, not just CLAUDE.md, so you can grade the file Cursor is metering.
So is there any case where a free model makes my config genuinely free?
Only a local model you run yourself (Ollama, llama.cpp) has a true marginal dollar cost of zero per token. Even then the config is not free: every token of CLAUDE.md you feed it occupies the context window and adds latency, because a local model still has to process the prefix on every turn. Smaller local models also have smaller context windows, so a 6,000-token config eats a larger share of what you have. Free of dollars is not free of tokens, and tokens are what the ccmd analyzer measures.
What is the one number I should look at?
Paste your file into ccmd.dev and read estimatedTokensFireEveryTurn. That is the count that fires on every turn regardless of model. On a paid model it is a dollar figure; on a free model it is the share of your message cap, free credit, or request quota that the file consumes before you have typed a word. Same number, different denominator. The fix is the same on both: cut the lines the analyzer flags as bloat, vague, or cache-busting.
Related on ccmd: cost is context, not model, CLAUDE.md and the weekly quota burn, the token cost audit.
Comments (••)
Leave a comment to see what others are saying.Public and anonymous. No signup.