The Consulting Firm Analogy
Imagine you run a consulting firm. You have two kinds of employees: fresh graduates who cost little but can handle large volumes of basic work - research, document organization, first drafts, process execution - and senior consultants who cost ten times more but solve hard problems: strategic judgment, key decisions, client negotiations.
The smart move is obvious: let the interns handle 80% of routine tasks, and bring in the consultants for the hard parts. This is Management 101 - but in AI, it has only recently been put into practice.
From "One Model Does Everything" to Division of Labor
In April 2026, this pattern got a name: the Advisor Pattern.
- Anthropic launched a beta on the Claude platform: Sonnet or Haiku executes, consulting Opus at key decision points. Official figures: +2.7 percentage points on SWE-bench Multilingual versus Sonnet alone, with roughly 12% lower cost.
- LangChain quickly shipped an open-source middleware implementation. Community reports are even more striking: Haiku+Opus doubled browsing-task scores while reducing cost; Sonnet+Opus improved SWE-bench results while saving money.
- Qwen Code v0.14.x productized the pattern with explicit, configurable "main model + lightweight model" orchestration, working flexibly with OpenRouter and OpenSpec.
- Claude Haiku: cheap, fast, sufficient for 80% of scenarios
- Claude Sonnet: mid-priced, strong general capability
- Claude Opus: expensive, slow, but most reliable on hard problems
- GPT-5.5: twice the price of 5.4, but noticeably better code quality
- Traditional approach: one Opus model does everything - reading, analysis, writing - because "what if some step needs deep reasoning?"
- Advisor approach: 1. Haiku reads the document and extracts key figures and sections (no world-class reasoning needed) 2. When a judgment like "does this accounting policy change constitute a material risk?" arises, Haiku packages the question and sends it to Opus 3. After Opus rules, Haiku continues - formatting, charts, email drafting 4. A final logic check of the report calls Opus again
- Confidence thresholds: the model scores its own certainty and escalates below a threshold
- Classifier routing: a small classifier first judges task type, then picks the model
- Feedback loops: post-hoc evaluation of the small model's decisions continuously tunes the escalation strategy
- Akshay's summary of the Advisor Pattern: https://substack.com/redirect/f2e9e328-077e-404c-b508-da55b2854c54
- LangChain DeepAgents Advisor middleware: https://substack.com/redirect/8dd7407b-1f1c-4a9d-920a-dd8edc053af2
- Anthropic Advisor strategy notes: https://substack.com/redirect/cabf8ccb-d0e4-4d42-9a31-66bffd28470f
- Qwen Code multi-model orchestration: https://substack.com/redirect/6c6b069a-4f0d-4c46-985e-e1970a93198d
This is not one company's clever idea - the industry is converging on this consensus.
Why Now?
The economic driver is simple: model capabilities are tiered, but prices are tiered much more steeply.
When a "good enough" cheap model can handle most tasks and only 10% of steps need an expensive model, using the expensive model everywhere is waste - like sending a principal architect to fix a printer.
The deeper reason: agent task complexity is rising. Early AI was single-turn Q&A. Today's agents run dozens of steps, call multiple tools, and process long contexts. Using the most expensive model at every step makes bills spiral out of control.
What Does the "Advisor" Actually Do?
Consider the task: "Analyze this financial report, identify risks, and give investment advice."
Result: Haiku does 90% of the work, Opus intervenes at only 2-3 key decision points. Total cost can be about one-fifth of the original, with quality nearly unchanged.
A Philosophical Question: How Does a Small Model Know It Can't Handle Something?
If a small model cannot correctly assess task difficulty, it may stay silent when it should ask for help, or ping the advisor unnecessarily - both destroy efficiency. Current solutions include:
Anthropic's Advisor tool and LangChain's DeepAgents middleware are, at their core, solving this meta-problem of "when to ask for help."
Practical Implications for Developers
1. Your "model stack" becomes as important as your tech stack. Not "GPT or Claude," but "which model version in which scenario, combined via which routing strategy." 2. Cost optimization has a new dimension. Beyond reducing tokens or using cheaper APIs, you can optimize via smart routing - cheap models do the heavy lifting, expensive models strike only where it matters. 3. Evaluation gets harder. In the multi-model era, you evaluate the *combination strategy* - at what task-allocation ratio is total cost lowest while quality stays acceptable?
This Is Not the End
Further ahead, models may learn self-routing - deciding internally whether a question needs deep thought or a quick answer, without external orchestration. Muse Spark's "compressed thinking" already hints at this capability.
Model boundaries may also blur: as open-source models get "good enough" and local deployment costs fall, the expensive/cheap distinction may rest not on parameter count but on deployment location (cloud vs local), response speed (real-time vs async), or specialization (general vs vertical).
Back to the Consulting Firm
The AI industry is waking from the illusion that "everyone is a full-stack consultant" and accepting a more pragmatic reality: division of labor beats going it alone. Cheap models are not inferior goods - they are the optimal solution for specific scenarios. Expensive models are not panaceas - they are insurance for critical decisions.
When the industry designs AI systems through the lens of *teams* rather than *individuals*, true large-scale application is only beginning.
---
Further reading