Anthropic's Cheat Code Almost Nobody Is Using: The Advisor Strategy

My context limit broke three times in one week while building a website.

First time: working on page load animations. Second time: running a site audit. Third time I didn't even know why — too many small code changes, and Claude was carrying the entire history of every edit in context from line one.

I upgraded from Pro to Claude Max. 90€ a month. Figured that was the end of it.

It wasn't.

That's when I started actually measuring what was consuming context — and found something Anthropic didn't exactly advertise, but that changes everything if you're building with AI agents.

It's called the Advisor Strategy.

What Is the Advisor Strategy?

The principle is brutally simple:

Haiku runs the entire task. The moment it hits something too complex — it calls Opus, gets the answer, and keeps going.

Opus doesn't work full-time. Opus is the consultant you call only when you need one.

Instead of running every line of CSS, every "thanks", every simple question through your most expensive model — Haiku carries the load, and Opus steps in only where it genuinely adds value.

Why It Works Better Than You'd Think

According to Anthropic's published benchmarks, this approach achieves double the performance — a jump from 19.7 to 41.2 — at 85% lower cost compared to running everything through Sonnet.

In real money: if your team or agents currently spend €500/month on Claude API calls, the same work with the Advisor Strategy can cost under €100.

The difference isn't worse results. The difference is you stop paying for Opus on tasks Haiku handles just as well.

What It Looks Like in Practice

Imagine a standard agent that receives a request, processes data, generates a response, and formats output. Without the Advisor Strategy: all 4 phases go through Opus. Expensive, slow, and unnecessary in 80% of cases.

With the Advisor Strategy: - Phases 2, 3, and 4 — Haiku (fast, cheap) - Phase 1 and critical decision points — Opus (only when needed)

How to Turn It On in Claude Code

This is where most tutorials overcomplicate things.

In Claude Code, there's no code. No scripts. No configuration files.

One command:

/model opusplan

That's it. Claude Code automatically uses Opus while you're planning (Plan mode) and switches to Sonnet for implementation. Opus burns tokens only when you're making architectural decisions — not while you're writing CSS or tweaking a button color.

Alternatively, set it at startup:

claude --model opusplan

From that point, every project you open runs on this principle — without you ever thinking about it again.

What I Gained After Implementing This

My context per session dropped from 61% to 38%.

Not because I'm doing less. Because the right model is doing the right job.

Animations, audits, small code changes — all Haiku now. Opus steps in for architectural decisions, complex debugging, or when someone needs to see the bigger picture.

Why Almost Nobody Uses This

Because it looks complicated until you see how simple it is.

One change in which model you use by default. One condition that asks "is this a Haiku task or an Opus task?" And suddenly your agents are both smarter and cheaper than the competition running everything through one expensive model.

The Bottom Line

If you're building agents, automated workflows, or just use Claude heavily day-to-day — the Advisor Strategy isn't an optimization. It's basic hygiene.

Pay for intelligence when you need it. Don't pay when you don't.