Fable Ran Out. The Advisor Didn't: Claude Code's Fable Advisor and Code Mode, Tested

I ran out of Fable on three accounts this week. Fable is the model I want for the hard parts, and it has its own weekly meter.

Then this tip went viral:

PSA: Using Fable as an advisor doesn’t touch your Fable weekly limit

— @akshdeeps_001 on X

The top reply said: “No, it’s deducted from your weekly.” Nobody posted numbers. So I tested it.

What the Advisor Is

The advisor is an experimental Claude Code feature. Your main model does the work. At key moments, such as before it commits to a plan, when an error keeps coming back, or before it calls a task done, it asks a stronger model for advice.

The advisor reads the whole conversation on every call, with no cache. So each call is expensive in tokens. The question is whose tokens.

The Test

Test 1: an account with Fable at 100% used. Extra usage was off, so nothing could bill. I ran Opus 5.5 with --advisor fable and forced one advisor call. It worked. The API record shows claude-fable-5-2, 41,571 input tokens.

Test 2: a fresh account, to see if the meter moves.

• Before: Fable weekly at 6%.
• The spend: ten Fable advisor calls on a 218K-token context. 2.18 million uncached input tokens, 67K output.
• After: Fable weekly at 6%.
• Calibration: an hour earlier, direct Fable use moved the same meter from 3% to 6%. That was 4 million tokens, mostly cached, so it cost about the same as the advisor spend.

So the advisor spend should have moved the meter about 3 points. It moved 0.

How to Use It

Run Opus as your main model and set Fable as the advisor:

/advisor fable

That saves advisorModel to your user settings. Opus does the routine turns. Fable checks the plan and the finish.

The tip is about the Fable meter, not your whole budget. The docs say so:

Subscription plans: advisor usage counts toward your plan’s usage limits

— Claude Code docs, advisor tool

The Other Lever: Code Mode

Claude Code 2.1.325 also ships Code Mode, off by default. Instead of one tool call per model turn, Claude writes JavaScript that makes many calls at once, then returns only what it needs to read.

I built a rough version of this a year ago: MCP tools as code APIs, for 96% less context. Now it is built in.

I ran four tasks on a 297-post repo, each with Code Mode on and off. Opus 5.5, advisor off. Every answer was correct, and the edits were identical in both modes.

• Multi-step work saves the most. The broken-links task made about 600 tool calls in two scripts. Up to 3x fewer input tokens.
• Even a one-line lookup saves about 30%. Code Mode starts with a 37K-token context instead of 54K.
• Speed was about the same. The time is in the model, not the tool calls.

What Code Mode Costs You

• No Chrome, computer use or Remote Control.
• No -p. It works only in the interactive terminal. My headless bench scripts can’t use it.
• No agent teams. It won’t start while they are on.
• Sessions are locked to it. A Code Mode session resumes only with Code Mode on.

Turn it on with CLAUDE_CODE_EXPERIMENTAL_CODE_MODE=1 in your settings env.

The Takeaway

When Fable runs out, you still have two levers. Code Mode stretches every token. The Fable advisor keeps Fable on the hard decisions after the meter says no.

Turn both on.