Skip to content

00. Model Selection & Vendor Strategy — The Five-Year-Old Version

Picking a model is like staffing a kitchen. Not every dish needs the master chef. Not every shift survives without one.


Imagine a busy restaurant.

The menu has a hundred items. Some dishes are simple — chop salad, pour soup, plate bread. Some dishes are complex — multi-course tasting menu, sauce reductions, garnish art. Some dishes are routine — daily curry, rice, dal, sabzi.

Now picture three kinds of cooks in the kitchen.

The master chef is brilliant. She thinks deeply. She invents. She can fix a broken sauce blind. But she is slow. And her hour costs as much as four normal hours.

The workhorse cook is steady. He has handled every standard dish a thousand times. He is fast. He is reliable. He costs one-quarter of the master chef per hour.

The apprentice is very fast. Very cheap. She is great at chopping. Great at plating. Great at simple labelled tasks. She cannot invent a new dish. She cannot rescue a burnt sauce.

Look. A good restaurant manager does not put the master chef on chopping onions. She also does not put the apprentice on the tasting menu. She matches the cook to the dish. That whole matching habit is what this module teaches. Simple, no?


A tiny kitchen picture

┌──────────────────────────────────────────────────┐
│                  TICKET COMES IN                 │
│                       │                          │
│        ┌──────────────┼──────────────┐           │
│        ▼              ▼              ▼           │
│   chop onions    daily curry   tasting menu      │
│        │              │              │           │
│        ▼              ▼              ▼           │
│   apprentice    workhorse cook   master chef     │
│   (small)       (mid-tier)       (frontier)      │
│        │              │              │           │
│        ▼              ▼              ▼           │
│   ₹0.01/dish   ₹0.30/dish      ₹3.00/dish        │
│   30 sec       2 minutes       10 minutes        │
└──────────────────────────────────────────────────┘

That picture is the whole module compressed. On the left, cheap and fast and narrow. On the right, expensive and slow and brilliant. The manager keeps moving tickets to the right kind of cook. That habit decides whether the kitchen makes money or bleeds it.


Five real kinds of tickets, five different cooks

Pretend we run an AI customer-support team. Watch the same incoming ticket pass through five kinds of cooks.

Ticket Best cook Why
"Is this email a refund request — yes or no?" apprentice Closed set, one decision. Master chef is overkill.
"Pull the order ID and amount from this email" workhorse Structured extraction. Mid-tier handles it cleanly.
"Draft a polite refund reply in Hindi and English" workhorse Generation in known shape. Mid-tier is plenty.
"Investigate why this customer's last 6 orders failed" master chef Multi-step reasoning, planning, tool use.
"Decide if this case needs a human escalation" master chef or a smaller model with a strict rubric Judgement call. Frontier sometimes earns its price here.

Three insights fall out of this table.

One — most tickets do not need the master chef. Two — the same kitchen needs all three cooks. Three — choosing the cook is a recurring decision, not a one-time setup. A mature team picks the cook per ticket, not per company.


The three placeholders you will see called back

Placeholder What it really is
the kitchen Your whole AI system — every step, every tool, every prompt
the cook A specific model tier — frontier, mid-tier, or small
the ticket One task the model must do — classify, extract, generate, plan, judge
the matching habit The reasoning by which you assign cook to ticket — cost, latency, quality, risk
the kitchen manager You, the applied AI engineer or lead, making that decision deliberately
the supplier The vendor behind the cook — Anthropic, OpenAI, Google, Meta, Mistral, you-on-GPU
the bake-off The structured taste test that picks the right cook for a ticket
the second supplier The backup vendor warm in the wings when the primary supplier breaks

We will keep returning to these names. Whenever a chapter says the cook, it means a model tier. Whenever a chapter says the supplier, it means a vendor. The metaphor stays steady so the ideas can move.


Why this module exists

Every applied AI interview opens with one of two questions.

"How would you pick a model for X?" "Your provider has a 4-hour outage — what happens?"

Both questions are about the same skill — staffing the kitchen and keeping it running when a supplier disappears.

Candidates who name models lose. Candidates who reason about picking models win. That reasoning is what this module trains.


What's coming

  1. 01-the-tier-anatomy.md — frontier, mid-tier, small, on-device. Who's actually who in 2026.

  2. 02-cost-latency-quality-triangle.md — you pick two. The third bends. Worked numbers.

  3. 03-bake-off-methodology.md — same eval set, same prompts, same judge. Statistical significance, not vibes.

  4. 04-task-to-tier-mapping.md — classification → small. Extraction → mid. Agent planning → frontier. The routing matrix.

  5. 05-closed-vs-open-weight.md — Anthropic, OpenAI, Google vs Llama, Mistral, Qwen, DeepSeek. Where each wins.

  6. 06-switching-cost-anatomy.md — prompt portability, tokenizer math, schema drift, tool-call format. The hidden tax.

  7. 07-dual-sourcing-fallback-chains.md — keeping a second supplier warm without paying twice.

  8. 08-rate-limits-and-quota.md — anatomy of rate limits. Negotiation. Headroom planning.

  9. 09-vendor-risk-and-outages.md — outages happen. Model EOLs happen. Plan accordingly.

  10. 10-model-upgrades-in-production.md — the playbook for swapping a model without breaking evals.

  11. 11-on-prem-vs-managed-economics.md — when self-hosting pays off. When it really does not.

  12. 12-pricing-anatomy.md — input vs output, cached vs uncached, batch vs realtime. Reading a vendor invoice.

  13. 13-honest-admission.md — what model selection still cannot answer cleanly.


Bridge. First we need to actually see the cooks. Frontier, mid-tier, small, on-device. Who is who in 2026, and what makes each one earn its place in the kitchen. → 01-the-tier-anatomy.md