Chapter 3: Claude Model Selection and Tradeoffs

1. The Claude Model Families

Opus for Maximum Capability

Opus is Anthropic's most capable tier — the senior specialist reserved for problems no one else can crack: multi-hour autonomous agents, complex refactors spanning 50+ files, financial modeling, and high-consequence research synthesis. It sustains logical coherence across long outputs (10,000+ tokens), but carries the highest latency, making it a poor fit for real-time chat. Its natural home is batch and asynchronous work. The current model is Claude Opus 4.8 (claude-opus-4-8), launched May 28, 2026, priced at $5.00 / $25.00 per million tokens (Fast Mode: $10 / $50).

Sonnet for Balanced Performance

Sonnet is the balanced middle tier and the recommended default for most production use cases — near-Opus quality at faster latency and lower cost. The current model, Claude Sonnet 4.6 (claude-sonnet-4-6), is priced at $3.00 / $15.00 per million tokens and ships a 1M-token context window at no surcharge. The practical guidance: start with Sonnet; move up to Opus only when you measure a quality gap, and down to Haiku only when you measure a cost or latency problem.

Haiku for Speed and Cost Efficiency

Haiku is the fastest and most cost-efficient tier, built for high-volume, latency-sensitive workloads: classification, routing, extraction, summarization, and moderation. Claude Haiku 4.5 (claude-haiku-4-5) is priced at $1.00 / $5.00 per million tokens, averages ~110 tokens/second (2–3x faster than Sonnet, 4–5x faster than Opus), and hits sub-1.5-second p95 latency for chat. Roughly 70–80% of typical production queries can be handled by Haiku.

The Comparison Matrix

2. Quality, Latency, and Cost Tradeoffs

Mapping Task Complexity to Model Tier

Worked example — a support pipeline uses all three tiers: (1) Haiku classifies the ticket in under 1.5s; (2) Sonnet drafts a routine billing reply over retrieved policy docs, its 1M window holding the corpus; (3) Opus handles an escalated technical case reasoning across full history, three log files, and a config diff. This is right-sizing — the better question is "which model for this step?"

Latency-Sensitive vs. Quality-Sensitive Workloads

Latency-sensitive workloads put a human in the loop, waiting — live chat, autocomplete, voice interfaces. Haiku's sub-1.5s p95 is a feature you can feel; Opus's high latency is disqualifying regardless of answer quality. Quality-sensitive workloads run in the background, where extra seconds cost nothing perceptible — overnight migrations, batch analysis, research synthesis. Here Opus shines. Name the dominant pressure first, then choose.

Adaptive Thinking Support by Model

Modern Claude models support adaptive thinking (thinking: {"type": "adaptive"}), where the model decides how much internal reasoning to perform, steered by output_config: {"effort": "low" | "medium" | "high" | "max"}. This is another lever on the triangle. The versioning wrinkle:

• On the newest models (Opus 4.8, Opus 4.7, Sonnet 5), the old fixed-budget syntax {"type": "enabled", "budget_tokens": N} is rejected with a 400 error — you must use adaptive thinking.
• On Opus 4.6 and Sonnet 4.6, fixed-budget is deprecated but still functional as a transitional escape hatch.
• On older models, fixed-budget is still correct (budget_tokens must be less than max_tokens).

3. Model Releases and Versioning

Model IDs Are Pinned Snapshots

Breaking Behavior Changes Across Releases

These changes are silent in source code and loud at runtime: deprecated parameters often remain in SDK request types, so code type-checks and compiles — but a value that worked on Sonnet 4.6 throws a 400 on Sonnet 5. A migration that only swaps the model-ID string will pass review and fail on the first live request.

The Model Lifecycle and Deprecation Policy

For publicly released models, customers with active deployments get at least 60 days' notice before retirement. Note: first-party retirement dates apply to Anthropic-operated platforms; Amazon Bedrock and Google Cloud set their own schedules.

4. Decision Frameworks for Model Choice

The Four Decision Dimensions

1. Task complexity — how much reasoning depth does the task genuinely require?
2. Latency sensitivity — is a human waiting on the response in real time?
3. Cost at scale — what does this cost when multiplied by traffic volume?
4. Output quality requirements — what is the cost of a wrong answer?

Cost Modeling Per Tier

Two cost features apply across all tiers: batch processing is 50% cheaper, and prompt caching cuts cached input cost by up to 90%. Worked example — extracting fields from 1M documents (2,000 in / 200 out each): on Opus ≈ $15,000; on Haiku ≈ $3,000 (5x saving, no quality loss); on Haiku + Batch API ≈ $1,500 — a 90% reduction, because the task never needed Opus. Tiered routing can cut costs 50–70% versus running everything through Opus.

Escalation and Fallback Patterns

The cascade pattern attempts the task with Haiku first, escalating to Sonnet or Opus only on low confidence or a failed validation. First-pass filtering runs Haiku as a cheap classifier, sending only qualified inputs to the more capable tiers. One caution: changing the model mid-conversation invalidates the prompt cache (caches are model-scoped) — a common workaround is to spawn a cheaper-model sub-agent for the sub-task while keeping the main loop pinned to one model.

Right-Sizing the Model to the Workload

The checklist: (1) default to Sonnet; (2) measure before escalating to Opus; (3) measure before dropping to Haiku (which serves 70–80% of queries); (4) route, don't guess; (5) layer in batch and caching. The anti-pattern to internalize is running everything through Opus "to be safe" — the costliest default there is.