Skip to main content

5.3 Model selection and tradeoffs

The current line-up​

ModelIDContextInput $/MTokOutput $/MTok
Claude Fable 5claude-fable-51M$10.00$50.00
Claude Opus 5claude-opus-51M$5.00$25.00
Claude Opus 4.8claude-opus-4-81M$5.00$25.00
Claude Sonnet 5claude-sonnet-51M$2.00$10.00
Claude Haiku 4.5claude-haiku-4-5200K$1.00$5.00

Use the exact ID strings — no date suffixes. Prices are first-party API rates; Bedrock and Vertex are partner-operated with separate pricing.

Choosing a tier​

TierFits
Opus (and Fable, above it)Hard reasoning, long-horizon agentic work, complex coding, ambiguous judgment, the orchestrating agent in a hierarchy
SonnetThe broad middle — production workloads that need strong quality at lower cost, high-volume extraction and analysis
HaikuHigh-volume, latency-sensitive, well-specified work: classification, routing, simple extraction, cheap subagents. Note the 200K window

Three tradeoff axes, and they do not move together:

AxisLever, cheapest first
CostPrompt caching → prune input → shorter output → batch → lower effort → smaller model
LatencyStreaming (time-to-first-token) → shorter output → lower effort → smaller model → fast mode on Opus 5/4.8
QualityBetter prompt and tool descriptions → higher effort → larger model → decomposition into passes
Measure the simpler alternative first

Before building a multi-model cascade, measure the more capable model at lower effort on the same tasks. Lower effort on a newer model often matches or beats a prior generation at high effort, and one model means one cache namespace — caches are model-scoped, so a cascade forfeits cache reuse across its models. Judge cost per completed task, not cost per request: a cheaper call that needs three retries is not cheaper.

Capability differences that change code​

CapabilityNote
Adaptive thinking{"type": "adaptive"} on all current models. On Claude Opus 5 thinking is on by default when thinking is omitted; on Opus 4.8/4.7 omitting it means no thinking. On Fable 5 it is always on and cannot be configured
Effort levelslow–max on current models; xhigh arrived with Opus 4.7
Sampling parametersRemoved on Opus 5 / 4.8 / 4.7, Sonnet 5, Fable 5 — 400 if sent
budget_tokensRemoved on those models; deprecated on the 4.6 family
Assistant prefillRemoved on 4.6 and later — 400 if sent. Use structured outputs instead
Fast modeClaude Opus 5 / Opus 4.8 only, first-party only, premium pricing
Context window1M on current models; 200K on Haiku 4.5, which also caps inline PDFs at 100 pages
Max outputUp to 128K on current models; streaming required for large values
Refusal handlingstop_reason: "refusal" with stop_details; server-side fallbacks can re-run the request on another model, first-party only
TokenizerChanged between generations — re-baseline token counts after migrating

Do not assert a capability from memory in production code: client.models.retrieve(id) reports max_input_tokens, max_tokens, and capabilities at runtime.

Breaking behaviour changes across releases​

A model upgrade is a breaking-change review, not a string swap.

Change classWhat to check
Removed parametersSampling parameters, budget_tokens, prefill — a request that worked last year now 400s
Changed defaultsThinking on/off when omitted; thinking display default; effort default
New stop reasonsrefusal must be handled before reading content
Tokenizer changeCost models, chunk sizes, and truncation thresholds all shift
Prompt sensitivityPrompts written for an older model are often too prescriptive and reduce quality; over-specified scaffolding should be deleted, not extended
Tool-call formattingJSON escaping in tool_use.input varies — parse, never string-match
Latency and turn lengthNewer models on hard tasks can run much longer per turn; timeouts, streaming, and progress UX need revisiting
Cache behaviourCaches are model-scoped: switching models starts cold

The migration sequence: run the eval set on the new model before editing prompts → fix hard API breaks → re-baseline tokens → re-tune effort → audit prompts for stale instructions → canary → cut over. Details on 2.2 Systems life cycle.

Quiz​

  1. Question 1

    A high-volume classification route runs on an Opus-tier model and its prompts reach 300K tokens. The team wants to move it to Claude Haiku 4.5 for cost. Which fact blocks the move as the route is currently written?

  2. Question 2

    Before building a two-model cascade that routes easy documents to a cheaper model, which measurement comes first?