Claude Model Selection & Tradeoffs.

Model selection is the API-level decision of which model ID to send to POST /v1/messages, balancing three axes: capability, latency, and cost per token. The production tiers are Opus (claude-opus-4-8) for complex agentic coding and enterprise work, Sonnet (claude-sonnet-5) for the best speed/intelligence balance, Haiku (claude-haiku-4-5) as the fastest, cheapest tier with near-frontier reasoning, and Fable 5 as the highest-capability, slower tier. Since Claude 4.6, model IDs are dateless but still pinned snapshots - not evergreen pointers, so a deployed app never silently changes model behavior underneath it.

Anthropic's recommended selection method is benchmark evals on your actual prompts and data, not intuition - start at the cheapest tier that plausibly covers the task, and upgrade only on a measured capability gap. Capabilities and token limits are also queryable programmatically via the Models API (max_input_tokens, max_tokens, capabilities), so a system can select or validate a model choice at run time rather than hard-coding assumptions.

Pricing and context. Per MTok (input/output): Opus 4.8 = $5/$25, Sonnet 5 = $3/$15 ($2/$10 introductory through 2026-08-31), Haiku 4.5 = $1/$5, Fable 5 = $10/$50. Context and output ceilings split the same way: Opus 4.8, Sonnet 5, and Fable 5 support a 1M-token context window with 128k max output; Haiku 4.5 is scoped to 200k context and 64k max output.

Latency ranks in the same order as capability, inverted: Haiku is fastest, then Sonnet, then Opus, then Fable slowest. This is the core tradeoff triangle the exam tests - you cannot maximize capability, latency, and cost simultaneously, so the architecture has to state which two matter more for the task at hand.

Adaptive thinking (thinking: {type: "adaptive"}) is supported on Opus 4.8, Sonnet 5, and Fable 5 (always-on on Fable); Haiku 4.5 does not support adaptive thinking, but supports extended thinking as an alternative. A task that genuinely needs adaptive thinking cannot be served by the cheapest tier - Sonnet 5 becomes the cheapest tier that qualifies.

The `effort` parameter (output_config.effort) trades intelligence for latency/cost within one model, and is often a better lever than switching tiers entirely. It defaults to high on Opus 4.8 and Sonnet 5; xhigh is recommended for most coding and agentic work. Tuning effort up or down inside a single model is cheaper to reason about and re-test than a full model migration.

Migrations are not drop-in swaps. Moving to Opus 4.7+, Sonnet 5, or Fable 5 removes manual extended thinking (thinking: {type: "enabled", budget_tokens: N} now returns a 400 - use adaptive thinking instead), rejects temperature/top_p/top_k sampling parameters, and 400s on assistant-turn prefills. Sonnet 5 also defaults adaptive thinking ON when thinking is omitted (Sonnet 4.6 ran thinking-off by default), and its newer tokenizer produces roughly 30% more tokens for identical text - shifting cost baselines even when the prompt hasn't changed. A separate research-preview fast mode gives up to 2.5x output speed at premium pricing on Opus 4.8/4.7 (Opus 4.7 fast mode is deprecated, removal 2026-07-24).