Claude Code Usage Limits: Cut the Cost of Every Task

Three levers, one session

Before you can cut the bill, know which control does what.

1. 1

   The strategy in one line

   The card at the top of the video states the trade in a single sentence: pair Opus as an advisor with Sonnet or Haiku as an executor, and get near Opus-level intelligence at a fraction of the cost. The rest of the recording is that sentence wired into a real session.

   

2. 2

   Sonnet 4.6 opens the session at zero context

   The session starts on Sonnet 4.6 under a Claude Pro plan, with the status line reading ctx 0% because nothing has been sent yet. Sonnet is the executor, so it handles the turns; the advisor is the model held in reserve.

   

3. 3

   Four commands that move the bill

   Typing /model opens the command list, and four entries matter for cost: /model sets the AI model (it reports Sonnet 4.6 at this point), /effort sets the effort level for model usage, /advisor configures the Advisor Tool, and /status reports version, model, account and API connectivity.

   

Set up the advisor

A slash command, a picker and a confirmation — under a minute of setup.

1. 4

   Type the advisor command

   /advisor is the entry point. Its inline description states the rule the whole page rests on: consult a stronger model for guidance at key moments during a task. The strong model is not invited to every turn.

   

2. 5

   Choose which model advises the executor

   The panel is explicit about the cost: the advisor runs server-side and uses additional tokens, and pairing Sonnet as the main model with Opus as the advisor gives near-Opus performance with reduced usage. Three options are on offer — Opus 4.6, Sonnet 4.6 and No advisor — with No advisor checked until you change it.

   

3. 6

   Advisor committed as Opus 4.6

   Confirming the pick leaves the line Advisor set to Opus 4.6 under the command. The prompt beside it is still empty and the status line still reads ctx 0%: nothing has been spent yet, and the executor is still Sonnet 4.6.

   

Advisor versus the old Opus-plan setup

Two ways to use a strong model without paying for it on every turn.

1. 7

   The rule the old setup followed

   The overlay names the habit the advisor replaces: planning goes to Opus, executing goes to Sonnet. That is one model at a time, swapped by hand — exactly the difference the next frame draws out.

   

2. 8

   The two setups side by side

   Both results sit in the same terminal. /advisor reports Advisor set to Opus 4.6; /model opusplan reports Set model to Opus 4.6 in plan mode, else Sonnet 4.6. In the video's own explanation, the advisor setup lets the executor call the advisor mid-task with a shared context, while Opus plan gives you one model at a time.

   

How the split actually works

One shared context, two models, and a tool call between them.

1. 9

   Executor every turn, advisor on demand

   The diagram makes the roles concrete. The Executor is Sonnet and runs every turn in the main loop; the Advisor is Opus and is reached through a tool call, on demand. The advisor never produces user-facing output — it only sends guidance back to the executor.

   

2. 10

   Why the two models share one context

   The circled box is the point of the feature: the Executor reads and writes a Shared context of conversation, tools and history, and the Advisor reviews that same context and sends advice back into it, so the hand-off needs no re-explaining of the task.

   

What the video's chart shows

Higher score at a lower cost per task — and the same shape over the API.

1. 11

   More score at a lower cost per task

   This is the video's own SWE-bench Multilingual chart, plotted against cost per agentic task. Sonnet 4.6 High with an Opus advisor sits at 74.8% for $0.96; Sonnet 4.6 High solo sits at 72.1% for $1.09. The creator reads that as roughly three points more for about twelve percent less, and says so in the narration.

   

2. 12

   The same split over the Message API

   The strategy is not Claude-Code-only. In the Message API example the executor is model: “claude-sonnet-4-6”, and the advisor arrives as a tool entry with type “advisor_20260301”, name “advisor”, model “claude-opus-4-6” and max_uses 3. The comment underneath is the reason to watch your meter: advisor tokens are reported separately in the usage block.

   

Keep an eye on the meter

Background watchers spend tokens too.

1. 13

   A watcher that is not free either

   Anthropic's Monitor Tool slide closes the video. The tool watches something in the background and reacts when it changes without pausing the main conversation, and the comparison table is blunt about the cost: the Monitor Tool consumes tokens when it triggers, while /loop consumes tokens with every execution.

   

Keep going