Anthropic just released its most capable AI yet, then published a guide on how to avoid using it. The reason is the most useful AI lesson of the year.
No. And the company that builds the most powerful AI on the market just proved it.
Here is the short answer to the question in the title, before you scroll another inch. You should not use your most powerful, most expensive AI model for every task. You should use it for the decisions that change the outcome, and hand the volume work to something cheaper and faster. The skill that decides which is which is not technical. It is judgment. And if you have run a business for any length of time, you already have more of it than you think.
That answer came into sharp focus this month. Anthropic released Fable 5, the strongest model it has ever shipped. Then its own developer team turned around and published two patterns for getting almost all of that quality without paying top-tier prices on every token. A company telling you to use less of its premium product is worth paying attention to. The reason it works is a principle you can use whether or not you ever touch a line of code.
Key Takeaways
- Using your most powerful AI model for every task is like paying a first-chair violinist to also park the cars. Expensive, and a waste of the talent.
- Anthropic’s two patterns, advisor and orchestrator, both reserve the premium model for judgment and hand the volume to a cheaper model.
- In Anthropic’s own testing, these hybrids held roughly 92 to 96 percent of the top model’s quality at 46 to 63 percent of the cost.
- The skill that makes this work is knowing when a decision matters. That is human judgment, and AI did not make it obsolete.
- You do not need an API to use this. You need to route judgment and volume on purpose instead of by accident.
Why Is Using Your Best AI for Everything Such an Expensive Mistake?
Let me tell you where I learned this the hard way, long before AI.
I spent years running a digital marketing agency. I stepped away from it in 2023. And the most expensive mistake I made in the early days was simple: I did the ten-dollar work and the ten-thousand-dollar work with the same level of my own attention. I formatted proposals. I answered routine email. I built spreadsheets. All while the decisions that actually moved the business, the positioning, the pricing, the hard client conversations, waited in line behind the busywork.
The problem was not effort. I had plenty of effort. The problem was that I was pointing my most valuable resource, my own judgment, at tasks that did not need it.
That is the exact trap people are walking into with AI right now.
The instinct is understandable. A new model comes out, everyone says it is the best, so you route everything through it. Every email. Every summary. Every draft. And the top model happily does all of it, at top-model prices. For Fable 5, the developer community reports it runs roughly five times the cost of a mid-tier model on the output work that drives most of the bill. Point it at every task in your day and you have taken a professional orchestra and hired the first-chair violinist to also sweep the stage and park the cars.
The talent is real. You are just spending it in the wrong places. And most people never notice, because the bill arrives quietly and the work still gets done. It just costs far more than it needed to, and your best thinking still ends up buried under volume.
What Did Anthropic Actually Prove?
Here is where it gets concrete, and where I want to be careful with the numbers, because made-up statistics are a personal insult to me and they should be to you too.
In early July 2026, Anthropic’s official developer account shared two patterns the team uses internally. Both split the work between a strong, expensive model and a cheaper, faster one. The strong model only steps in at the moments that matter. The cheaper model carries the load.
Pattern one is the advisor. Picture a senior consultant you keep on retainer. You do not pay that person to answer your email. You pay them for the two or three moments a month where a wrong call costs real money. In this pattern, a cheaper model (Sonnet 5) does the whole job start to finish, and only consults the premium model (Fable 5) when it hits a genuine decision point. In Anthropic’s testing on a coding benchmark called SWE-bench Pro, this combination reached about 92 percent of the top model’s solo score at about 63 percent of the cost. The premium model was called, on average, about once per task. Once.
Pattern two is the orchestrator. Picture a good general contractor on a job site. The contractor does not swing every hammer. The contractor reads the blueprint, decides what gets built and in what order, assigns the crews, and checks the work. In this pattern, the premium model plans, breaks the job into pieces, hands those pieces to cheaper worker models, and assembles the result. On a research benchmark called BrowseComp, Anthropic reported this reached about 96 percent of the top model’s score at about 46 percent of the cost.
Now the honesty part. Those are Anthropic’s own internal test results, on two specific benchmark tasks, not independent studies. One of the features involved is still in beta. So treat the numbers as directional, not as a promise for your exact work. Your results will move based on your tasks. What does not move is the principle underneath them, and the principle is the whole point.
The Real Lesson: Your Judgment Is the Routing Layer
Here is the part almost everyone reading the news about this missed.
The best tool in the building should not do every job in the building. And the thing that decides which jobs deserve the expensive brain is judgment.
Sit with that, because it is the exact opposite of the story you have been sold. You have been told AI made your experience worth less. That the machine does the thinking now, so the twenty years of instinct you built is a depreciating asset. I hear this fear constantly from the community I teach, and it has a name. I call it confidence collapse. It is not that people are overwhelmed by AI. It is that they secretly believe their hard-won expertise just became worthless.
Anthropic just published the counterexample. Their entire cost strategy depends on something knowing when a decision matters. In their system, that something is another model. In your business, that something is you.
Your experience is not obsolete. Your experience is the routing layer. It is the instinct that looks at a task and knows, in about two seconds, whether it needs your full attention or whether it can run on autopilot. A beginner cannot do that. A beginner escalates everything, or nothing. You escalate the right things. That instinct is the asset, and no model replaces it. It is the whole reason I teach that you are the conductor, and AI is the orchestra. The conductor never plays an instrument. The conductor decides who plays, and when.
How Do You Apply Premium-for-Judgment in a Business That Doesn’t Write Code?
You are probably not going to wire two AI models together through an API. Good. The lesson was never about the plumbing.
The lesson is about doing on purpose what most people do by accident. When you use AI, you are already making routing decisions. You are deciding when to think hard yourself, when to let the tool run, and when to double-check its work. Most people make those calls unconsciously and inconsistently. The upgrade is to make them deliberately.
Here is what that looked like when I stopped doing the ten-dollar work myself, translated for AI.
I used to sit down and ask my best AI tool to grind out forty social captions for the week. Slow, expensive, and honestly the output was fine but forgettable. Now I use the strong model for one job only: locking the angle, the hook, and the single idea worth repeating. That is the judgment work. Once that is set, the volume, the forty variations, the summaries, the repurposed versions, runs on a faster, cheaper pass against that plan. I reserve the premium thinking for the decision that shapes all forty, not for the forty individually.
This is the same principle I teach as one agent, many workflows. You are not buying five tools and juggling them. You are running one command center and routing the work inside it. Premium intelligence for the decisions that compound. Cheap, fast execution for the volume. Your judgment sitting in the middle, deciding which is which.
It mirrors good delegation exactly, which is why experienced operators pick it up fast. When you hand a task to a team member, you already ask yourself: does this need me, or can someone else run with it, and where do I check in? You have been doing premium-for-judgment routing with humans for years. AI just gave you a faster, cheaper set of players to route to.
How to Route Your AI Work This Week (Step by Step)
You can put this to work before Friday. Here is the sequence.
- List the AI-assisted tasks you actually do. Write down the eight to twelve things you use AI for in a normal week. Content, email, research, summaries, proposals, customer replies, whatever they are. You cannot route what you have not named.
- Sort each task into judgment or volume. For each one, ask a single question: does the outcome change based on how good the thinking is, or is the outcome basically the same no matter what? Strategy, positioning, and hard decisions are judgment. Drafting variations, cleanup, and formatting are volume.
- Assign the judgment tasks to your strongest tool or your own attention. These are the moments that deserve the premium. Do not rush them and do not automate them into the ground. This is where quality is decided.
- Assign the volume tasks to a faster, cheaper pass. Most AI platforms let you pick a lighter, faster model or mode. Use it for the bulk work. If you are on a single tool, this can simply mean a quick, low-effort pass rather than a slow, deliberate one.
- Set your escalation trigger. Decide, in advance, what makes a volume task jump back up to judgment. A high-stakes client. A number that has to be right. A public-facing claim. When the trigger fires, escalate. When it does not, let the cheap lane run.
- Add one checkpoint. Before anything ships, do a single review pass with your judgment layer, you or your strong tool. One good check catches most of the errors a cheaper pass makes before they reach a customer.
- Review the map monthly. Your business changes. A task that was volume last quarter might become judgment when the stakes rise. Spend ten minutes a month re-sorting. The routing map is a living thing, not a one-time setup.
Frequently Asked Questions
Do I need to use Claude or a specific tool to apply this?
No. The advisor and orchestrator patterns came from Anthropic, but the principle works on any AI platform. Use your strongest model or mode for judgment and decisions, and a faster, cheaper one for volume work. The tool matters far less than the routing habit you build around it.
Isn’t the cheaper model just going to produce worse results?
Not for the right tasks. Cheaper models are highly capable at execution, drafting, and volume. They struggle more with complex judgment. The whole point is to match the model to the task, so the cheaper one handles work where quality is already locked in by your plan, not work where deep reasoning decides the outcome.
How do I know when a task deserves the premium model?
Ask whether better thinking changes the result. If a smarter approach meaningfully improves the outcome, route it to your strongest tool. If the outcome is essentially the same regardless, route it to the cheap, fast lane. Over time this becomes a two-second instinct, the same one you use when deciding what to delegate to your team.
Does hybrid routing always beat just using the cheap model?
No, and it is worth saying plainly. If a task is simple and the plan is obvious, the cheaper model alone is often the right call, with no premium step needed. Routing adds value only when premium judgment actually improves the result. Do not add complexity where a single fast pass already wins.
Did AI just make my years of experience worthless?
The opposite. Anthropic’s entire cost strategy depends on something knowing when a decision matters, and in your business that something is you. Your experience is the routing layer, the instinct that decides where premium thinking earns its keep. Beginners cannot route well. Experienced operators can. That is your advantage.
The Skill Nobody Is Selling You
Go back to where we started. The most powerful AI company on the market built its best model, then taught everyone how to mostly avoid using it. Not because the model is bad. Because the smartest way to use any powerful, expensive resource is to reserve it for the moments that deserve it.
That is not a story about AI getting smarter. It is a story about a skill getting more valuable, and the skill is yours. Knowing when the decision matters. Knowing when to spend the premium and when to let the cheap lane run. Knowing which two seconds of the week actually change the outcome.
You have been building that instinct your entire career. AI did not erase it. AI just made it the most important thing you bring to the table. The people who win the next two years will not be the ones with the most expensive model. They will be the ones who know when to spend it.
You are the conductor. You always were. Now go decide who plays.
P.S. Inside AI Insiders, we turn these routing decisions into repeatable systems, so your best thinking lands where it compounds and everything else runs without you. If that is how you want to operate, the door is here. [LINK: AI Insiders]
Author Bio
Jonathan Mast is the founder of White Beard Strategies and an AI educator who teaches practical AI adoption to entrepreneurs, coaches, and consultants. He built and exited a seven-figure digital marketing agency, and now leads a community of more than 500,000 people learning to use AI to amplify their expertise rather than replace it. His core message is simple: you are the conductor, and AI is the orchestra.
