Choosing the Right Claude Model

Picking the wrong Claude model for a task is one of the most common and expensive mistakes in production AI systems. Using Opus for simple classification burns budget. Using Haiku for nuanced legal analysis produces worse results than you'd get with a better model. Getting model selection right is as important as getting the prompt right.

The Claude Model Family

Anthropic offers four tiers: Haiku (fast/light), Sonnet (balanced), Opus (long-running agentic and knowledge work), and Fable (the top tier, for the most demanding reasoning and long-horizon agentic work).

Example model IDs (always check the models overview for the current list):

Model	ID	Context	Relative Cost
Claude Haiku 5.5	claude-haiku-5-5	1M	Lowest
Claude Sonnet 5.5	claude-sonnet-5-5	1M	Mid
Claude Opus 5.5	claude-opus-5-5	1M	High
Claude Fable 5.2	claude-fable-5-2	1M	Highest

Always verify current model IDs and pricing in the models overview — Anthropic updates the model lineup regularly and pricing changes with each generation.

The generation number matters: a newer-generation Sonnet usually beats an older-generation Opus. Don't assume "Opus" automatically means the best available model.

Cost, Speed, and Quality Trade-offs

These three attributes form a triangle — you typically optimize for two at the expense of the third.

Haiku is optimized for cost and speed:

• Fastest of the tiers
• Lowest input and output token cost (a small fraction of the Opus price per token)
• Strong on structured tasks, classification, simple transformations
• Weaker on multi-step reasoning, nuanced judgment, complex writing

Sonnet is the workhorse — the balanced choice that handles the vast majority of real-world tasks:

• Significantly better reasoning than Haiku
• Much faster and cheaper than Opus
• The right default when you haven't profiled the task yet
• Handles most writing, analysis, coding, and summarization tasks well

Opus is for tasks where quality matters more than cost:

• Stronger reasoning than Sonnet, especially for tasks requiring sustained multi-step logic
• Strong at following complex, nuanced instructions
• Slower latency (usually 2-5x Sonnet's time-to-first-token for long outputs)
• Costs significantly more per token
• For interactive applications with latency requirements, often Sonnet wins anyway

Fable sits above Opus:

• Built for the most demanding reasoning and long-horizon agentic work
• The most expensive tier per token
• Worth testing when Opus, even at a higher effort setting, still falls short on your task

Anthropic's own guidance for people who are unsure is to start with the current Opus for most workloads and move up to Fable only when Opus is not enough. The cost-first workflow in this lesson starts lower and moves up. Both work. Pick based on whether cost or quality is your tighter constraint.

The answer: steps 1 and 2 (extraction and classification) are good candidates for Haiku. Step 3 (flagging unusual clauses) requires legal judgment — Sonnet at minimum, Opus if the stakes are high.

When to Use Each Model

Haiku: High-Volume, Structured Tasks

Haiku shines when you need to process many requests quickly and cheaply, and the task is well-defined.

Good fits for Haiku:

• Intent classification ("is this message a complaint, question, or compliment?")
• Entity extraction from structured text
• Language detection
• Sentiment scoring
• Simple summarization (bullet points from a meeting transcript)
• Routing decisions in multi-agent pipelines
• Generating short, templated outputs (subject lines, tags, labels)

The Haiku prompting implication: Haiku needs more explicit structure. Because it has less reasoning capacity, ambiguity is more costly. Your prompts for Haiku should be:

• More explicit about output format
• Simpler in structure (fewer competing instructions)
• Heavier on examples (few-shot helps Haiku more than Opus)

Sonnet: The Default for Most Tasks

When in doubt, start with Sonnet. It handles the vast majority of real-world tasks at reasonable cost and speed.

Good fits for Sonnet:

• General coding assistance and code review
• Document summarization and analysis
• Email drafting and editing
• Customer support response generation
• Research synthesis
• Reasonably complex data extraction
• Most RAG (retrieval-augmented generation) pipelines

The Sonnet prompting implication: Sonnet is capable enough that you don't need to over-specify. Give it clear context and constraints, but you don't need to hand-hold through every step. It can handle moderate ambiguity and recover from slightly imprecise instructions.

Opus: Complex Reasoning and High-Stakes Tasks

Reserve Opus for tasks where quality is genuinely critical and you need the model's best judgment.

Good fits for Opus:

• Complex legal or financial document analysis
• Multi-step mathematical or scientific reasoning
• Code architecture design and system design review
• Tasks requiring sustained chain-of-thought over many steps
• Evaluating other models' outputs (LLM-as-judge)
• Creative work where quality differentiation matters
• Tasks where you're prompting Claude to reason about its own reasoning

Use Fable for the same kinds of tasks when they are harder or run longer: demanding reasoning and long agentic work where Opus still falls short. The prompting advice for Opus below applies to Fable too.

The Opus prompting implication: Opus can handle more complexity in a single prompt. You can give it less scaffolding and more latitude. Opus is better at inferring your intent from minimal instructions. That said, clear prompts still outperform vague ones — being brief is not the same as being unclear.

How Model Choice Affects Prompting Strategy

This is the part most guides skip, and it matters a lot.

With Haiku, use more structure:

With Sonnet, moderate structure:

With Opus, you can be more open-ended:

Same underlying task, three different levels of scaffolding. Haiku needs rails. Opus can navigate open terrain.

Token Pricing: Thinking in Costs

When you're running prompts at scale, model cost becomes a real factor. While exact pricing changes, the relative structure is stable:

Tier	Typical relative cost (per token)
Haiku	1x (baseline)
Sonnet	roughly 20x Haiku
Opus	roughly 40x Haiku
Fable	roughly 80x Haiku

A useful mental model for the current lineup: the jump from Haiku to Sonnet is the big one. After that, each tier up costs about twice the one below it. These ratios shift with each model generation, so check the pricing page before you budget.

The ROI question: Is the quality improvement from Sonnet worth roughly 20x the cost for this task? Is the quality improvement from Opus worth about double the Sonnet price? For classification, usually no. For contract review, often yes.

Decision Framework

Use this checklist when selecting a model:

1. Is this a high-volume task (>1000 requests/day)?
   YES → Start with Haiku, benchmark quality, upgrade if needed
   NO  → Move to step 2

2. Does the task require multi-step reasoning or nuanced judgment?
   NO  → Sonnet is probably sufficient
   YES → Move to step 3

3. Are the stakes high (legal, financial, medical, safety-critical)?
   NO  → Sonnet is probably sufficient
   YES → Consider Opus, or Fable for the hardest problems

4. Is latency a hard constraint?
   YES → Haiku or Sonnet (Opus adds latency)
   NO  → Let quality drive the decision

A good default workflow: prototype with Sonnet, optimize with Haiku where cost matters, upgrade to Opus only for tasks where Sonnet's output is genuinely insufficient.

Exercise: Design a Model-Appropriate Prompt

Key Takeaways

• Haiku for high-volume, structured tasks — use more scaffolding, explicit format constraints, and few-shot examples
• Sonnet for most tasks — the reliable default that balances cost, speed, and quality
• Opus for complex reasoning and high-stakes tasks where quality justifies the cost
• Fable is the top tier, for the most demanding reasoning and long-horizon agentic work when Opus falls short
• A newer-generation Sonnet often outperforms an older-generation Opus — check release dates, not just tier names
• Model choice directly affects prompting strategy: Haiku needs more structure, Opus can handle more ambiguity
• Profile your tasks before committing to a model — quality differences vary significantly by task type