The thing nobody tells you when you start working with AI agents seriously is that model selection is an engineering decision, not a philosophical one. Pick wrong and you're either burning money on a model that's solving the wrong class of problem, or you're watching a cheap model fail on something that needed proper reasoning, and assuming the whole approach doesn't work.
It does work. You just reached for the wrong tool.
A framework that actually helps
I think about model selection across four axes:
- Task structure — is this well-defined or ambiguous?
- Context requirement — how much does the model need to hold in its head at once?
- Reasoning depth — does it need to plan, or just retrieve and format?
- Error cost — what happens if it gets it wrong?
Map any task to those four dimensions and the model choice mostly selects itself.
The tiers (as of mid-2026)
Tier 1: Frontier reasoning (Claude Opus, GPT-4o, Gemini Ultra)
These are for ambiguous, multi-step problems where intermediate reasoning determines correctness. Code refactors that span multiple files. Architectural design questions. Anything where a wrong intermediate step cascades into a structurally broken output.
Cost: significant. Use sparingly and deliberately.
Tier 2: Fast, capable (Claude Sonnet, GPT-4o-mini)
The workhorse tier. Well-structured tasks with clear inputs and outputs. Writing a test suite for known function signatures. Summarising a document into a fixed schema. Drafting copy from a detailed brief. This is where 80% of your actual agent work lives.
Cost: 5–15× cheaper than Tier 1. Use by default.
Tier 3: High-throughput, low-latency (Claude Haiku, GPT-3.5-turbo class)
Classification, routing, slot-filling, extraction from structured text. If the task is "read this input, output one of these five labels", a frontier model is overkill. Haiku will do it faster and for a fraction of the cost.
Cost: 20–50× cheaper than Tier 1. Use for high-volume repetitive subtasks.
The decision that catches people out
The mistake I see most often — including in my own early pipelines — is using a Tier 1 model to handle tasks that belong to Tier 3, because Tier 1 always succeeds and Tier 3 sometimes doesn't.
That's backwards reasoning. The correct response to a Tier 3 model failing on a task is to improve the prompt and the task specification, not to upgrade the model. A well-structured classification prompt that gives Haiku explicit output format constraints, examples, and a fallback category will outperform a vague Opus prompt on the same task — and cost 40× less.
If you can't make a cheaper model succeed with a better prompt, you may genuinely have a Tier 2 or Tier 1 task. But verify that first.
Context window is a different problem
Context window size and model capability are not the same dimension. A 200k context window doesn't mean you need a frontier model — it means you have a lot of input material. Haiku has a large context window. Use it for RAG retrieval tasks where you're providing a big document and asking for extraction. Use Opus only if the extraction requires sophisticated cross-document reasoning.
In practice, most RAG pipelines are retrieval + formatting tasks. Tier 2 or Tier 3. The embeddings are doing the heavy work. The model is doing finishing work.
My personal routing logic
For lifedrawing.art's n8n-based AI pipelines:
- Incoming question classification → Haiku
- Answer generation from retrieved FAQ content → Sonnet
- Evaluating whether to escalate to human → Sonnet (error cost is high — false confidence should escalate, not fail silently)
- Drafting the escalation email summary → Sonnet
Opus appears nowhere in that pipeline. The task is well-structured and the inputs are known.
The actual cost calculation
Before you pick a model, do the maths. Sonnet at $3/million input tokens vs Opus at $15/million. If your agent handles 10,000 requests per month at an average of 2,000 tokens each:
- Sonnet: £47/month
- Opus: £237/month
For many tasks, Sonnet and Opus produce equivalent outputs. That's £190/month in capability you're not using.
The right question is not "which model is best?" It's "which model is sufficient for this task, at this error rate, at this volume?"
Answer that honestly and the decision is mostly arithmetic.
These observations come from running production AI pipelines at lifedrawing.art and from the AI tooling work I presented at the News UK All Hands Engineering Meeting in early 2025.