Select appropriate Claude models based on trade-offs

Key points

1. 1

   Think in tiers and trade-offs, not names or prices. The current families are Claude 5 (Fable 5.1 as the most capable model for the hardest reasoning and long-horizon agentic work, Opus 5.5 in the Opus tier, Sonnet 5 as the balanced tier) plus Haiku 4.5 as the small, fast, low-cost tier. Higher tiers buy capability at the cost of latency and per-token price.

2. 2

   Start from a quality bar measured on your own representative data, then pick the smallest model that meets it. A quality gap on a labelled sample is evidence; 'the text is nuanced' is not.

3. 3

   The small tier fits high-volume, low-stakes, latency-bound work: classification, routing, extraction with clear rules, pre-screening of inputs, and worker or subagent roles under a stronger orchestrator.

4. 4

   Escalate to a higher tier when an eval gap cannot be closed with prompt tuning and examples, when the task needs reasoning across several documents, or when it is a long-horizon agentic loop where small per-step errors compound into failed runs.

5. 5

   Route by task for mixed workloads: a lightweight classifier or rules sends simple requests to the fast tier and hard ones to a capable tier. One model for everything either overspends on the easy majority or underserves the hard minority.

6. 6

   Symptoms of model mismatch after the prompt has been tuned: looping, re-reading the same files, stopping short on multi-step work, missing conditions that interact across documents. More turns, more emphasis or a reviewer of the same tier do not fix these.

7. 7

   'Switch to a bigger model' is the exam's favourite non-answer for problems that live elsewhere: vague prompts, missing programmatic guardrails, broken retrieval, bloated tool sets, or a need for determinism. A larger model does not provide reproducibility or authorisation.

8. 8

   Before building a two-model cascade, measure the more capable model at a lower effort setting on the same tasks; it often meets the bar at comparable cost with far less machinery. A single model means one prompt cache namespace, one thing to monitor and no confidence threshold to tune.

9. 9

   Judge cost per completed task, not per request. A cheaper model that needs more retries, turns or human corrections is not cheaper.

10. 10

    Current models use adaptive thinking with effort controlling depth; the fixed thinking budget belongs to older models such as Haiku 4.5. Do not build keys on which sampling parameters a specific model accepts.

11. 11

    Context window size is a per-model property, not a tier ranking; check the model's specification rather than assuming a larger tier always has a larger window.

12. 12

    Pin the model per route or per conversation. Switching models mid-conversation invalidates the prompt cache and changes behaviour, and prompts tuned for one model may over- or under-trigger tools on another, so any model change is re-evaluated.

13. 13

    When presenting the choice to stakeholders, state the measured trade-off: quality on the eval, time-to-first-token and total latency, cost per task, and any feature or platform availability constraints.

Read the source

• Models overview
• Choosing a model
• Effort
• Building effective agents