
Which Claude model should you build with? Fable, Opus, Sonnet, or Haiku

The article guides model selection across the current Claude lineup: Claude Haiku 5.5, Claude Sonnet 5.5, Claude Opus 5.5, and Claude Fable 5.2, with Claude Mythos 5.2 reserved for cybersecurity and life sciences research via verification programs. Start most work on Opus 5.5 at medium effort, Claude Code‘s default. Move down to Sonnet 5.5 for well-scoped tasks and Haiku 5.5 for high-volume classification, routing, extraction, and subagent work; move up to Fable 5.2 for the hardest technical projects or when Opus 5.5 at high falls short. Adjust effort before switching models: one level down can suffice, and one level past high can cost as much as a bigger model. Higher effort adds latency but does not add knowledge or skill the model lacks.
Side by side, Haiku 5.5 is fastest and cheapest at $0.10/$0.50 per million tokens up to 100K and $0.50/$2.50 above; Sonnet 5.5 is $2/$10; Opus 5.5 is $4/$20; Fable 5.2 is $8/$40. All have 1M context and 128K max output; Opus and Fable add fast mode. Opus and Fable always think, Sonnet thinks by default with a between_tools option, and Haiku thinks by default but can turn thinking off at high effort or below.
Trade-offs: higher per-token price is not always higher per task. On Terminal-Bench 3.0, Fable 5.1 at max scored 58.0%, while Opus 5.5 at high scored 58.9% with half the tokens. On FrontierCode 2.0, Fable 5.2 at high scored 45.3% versus Opus 5.5 at max 45.5%, for roughly half the cost per task. Hardest tasks drive spend: in one 20-problem search benchmark, two problems carried 43% of spend. On SWE-bench Pro, Opus 5.5 at medium scored about 2.5 points lower for about 70% of high’s cost; at low about 8 points lower for a third; at xhigh about 1.4 points higher for 2.5 times the cost. Prefer self-verification, then one effort level up, then the next model at lower effort. Opus 5.5 at low with failures re-run at high passed about 97% of tasks for $0.17 each versus 95.3% for $0.29 all at high.
Context costs on every turn, though long runs mostly bill as cache reads. Opus and Fable charge $0.20 per million cache reads, Sonnet $0.10; a turn’s cache read costs about $0.03 at 150K tokens and $0.004 at 20K. Keep the expensive model’s context lean and let cheaper subagents read verbose material. Haiku’s price varies by prompt length: under 100K at $0.10 input, over 100K at five times that rate, still a quarter of Sonnet’s per-token price; for long prompts, compare Haiku and Sonnet on cost per finished task.
When mixing, switch at a natural break: prompt cache belongs to the model that wrote it, and the first turn on a new model pays the cache-write price, so run /compact or start fresh. Thinking blocks are selectively readable: Opus 5.5 reads Sonnet and Haiku blocks, Sonnet reads Haiku blocks, Fable 5.2 reads Opus, Sonnet, and Haiku blocks; moving down a tier loses reasoning. Subagents bill at their model’s rates; set model: haiku or model: sonnet in a subagent definition or set CLAUDE_CODE_SUBAGENT_MODEL. Keep small models on cheap-to-spot mistakes and code edits on the main model. Claude Code‘s built-in Explore subagent currently runs on the session model; define your own with model: haiku. The haiku alias resolves to Haiku 5.5 on the Anthropic API with Claude Code v2.1.293+ but to Haiku 4.5 on Bedrock, Google Cloud, Microsoft Foundry, and Claude Platform on AWS. The opusplan alias plans with Opus and executes with Sonnet; measure it first.
The advisor tool (beta) pairs a cheaper executor with a stronger advisor, billing most tokens at executor rates; the cost guide cautions none of four pairings clearly beat the executor alone. One internal coding benchmark, Opus 5.5 executor at high with Fable 5.1 advisor, scored 90.1% at $2.92, 1.7 points over Opus alone for about 2.1 times the money. For work larger than one context window, a top-tier lead can split tasks among cheaper parallel workers; an internal defect-search benchmark over about 22 million tokens with Opus 5.5 leading Sonnet workers cost about a third less than Opus alone at default effort with similar score. Skip multiagent orchestration for single chains or one-context work.
Compatibility matters: Fable 5.2 extends preserved thinking to every API account, and the API returns a 400 error when anything before a Fable 5.2 thinking block changes unless you opt into dropping it; keep history append-only. tool_choice any or tool returns a 400 on Sonnet, Opus, and Fable; use auto, mark tools strict, and say when to use them. Computer use works only through computer_toolset_20260801 on the Claude API and Google Cloud managed. Migrating Haiku 4.5 to 5.5 means replacing budget_tokens with effort, dropping non-default temperature/top_p/top_k, and ending with a user turn. Safety classifiers can fall back: flagged cybersecurity to Opus 4.8, biology to Opus 5. Build requests for the strictest model when running several.
Finally, measure on your own work. Use published results only as a starting point; sample real tasks, write outcome checks, run candidates at two or three effort levels in separate sessions, record per-task cost from the usage field, plot score against spend, and keep a multi-model setup only if it beats every single-model point. Run the winner in shadow first; in Claude Code, /model and /usage give a quick check.

