Which Claude Model Should I Use?
Answer three quick questions and get a model-tier recommendation with the reasoning.
Answer all three to see a recommendation.
What this tool does
This picker suggests which Claude model tier to start with: Haiku, Sonnet or Opus. You answer three multiple-choice questions about the task, what matters most, and the expected volume. It names one tier, gives a short reason, and says when to move up or down.
It is meant for developers and technical leads who are about to wire a model into an application and want a sensible default before testing. It does not look at your prompts or data, and it does not measure quality or cost, so treat the result as a first try to confirm on your own inputs.
How it works
Each answer is scored 0, 1 or 2 by its position in the list (first option 0, last option 2). Task and priority count double, one volume answer subtracts a point, and the total is compared with two fixed thresholds. The scoring is a few lines of arithmetic in the page, and the tool makes no network request for it.
- Points
- pts = task * 2 + priority * 2
- High-volume nudge
- if volume = 'Very high throughput': pts = pts - 1
- Recommendation
- pts <= 2: Haiku; pts >= 6: Opus; otherwise Sonnet
What it assumes
- The weights are a judgement call, not a measured result.
- Only 'Very high throughput' changes the score. 'Normal' and 'Low / occasional' give the same result, and the nudge can only make Opus harder to reach; it never changes a Haiku result.
- The tier names shown (Claude Haiku 4.5, Sonnet 4.6, Opus 4.8) are labels in the code and will drift as new models ship. The tool carries no prices or context-window sizes and does not consider latency targets, tool use, compliance or data residency.
Worked example
An illustrative case: a team is building a general-purpose assistant, wants the best quality it can get, and expects very high request volume.
| Input | Value |
|---|---|
| What's the task? | General: chat, drafting, everyday coding (score 1) |
| What matters most? | Best possible quality (score 2) |
| Expected volume? | Very high throughput |
- pts = 1 * 2 + 2 * 2 = 6
- Very high throughput, so pts = 6 - 1 = 5
- 5 is neither <= 2 nor >= 6, so the tier is Sonnet
- For comparison, 'Normal' or 'Low / occasional' volume leaves pts at 6, which gives Opus
Recommended: Claude Sonnet 4.6 (Balanced default), with the note 'For the hardest 10% of inputs, route just those to Opus; for cheap bulk work, drop to Haiku.'
Frequently asked questions
- Is the recommendation based on benchmarks or my real data?
- No. It is a rule of thumb from three answers and a fixed scoring rule. Treat it as the tier to try first, then compare tiers on a sample of your own inputs.
- Why do I sometimes get Sonnet when I pick high-quality options?
- Opus needs 6 or more points. At normal or low volume the task and priority scores only need to add up to 3. With 'Very high throughput' they must add up to 4, meaning both 'Hard' and 'Best possible quality'.
- Does the tool know current prices or context-window limits?
- No. It shows a tier name, a one-line description and a suggestion for moving up or down. Check Anthropic's current pricing and model documentation for exact costs, limits and version names.
- What does escalating to Opus on the hard cases mean in practice?
- Run most requests on a cheaper tier and send only the difficult ones to a more capable model, for example after a failed validation or a low-confidence answer. If only a small share of requests needs the stronger model, average cost stays closer to the cheaper tier; the tool does not estimate that share.
- What should I do after getting a recommendation?
- Build a small test set of real inputs and run it on the recommended tier and the one on each side. Compare quality, latency and cost, then pick the cheapest tier that meets your quality bar.
Read the guide
Building this for real?
I help teams design and ship AI systems — cost modelling, RAG, agents, and architecture.
Work with me