Fable, Opus, Sonnet or Haiku: Key Takeaways from Claude's Model Guide

Key takeaways from Anthropic's November 2026 model guide: start on Opus 5.5, tune effort before switching models, pick by task, and run subagents on smaller models. Includes official prices for all four models.

Short answer: if unsure, use Opus 5.5 at its default medium effort; if results fall short, raise effort first, and only then move to a stronger model. This is a summary of Anthropic's November 9, 2026 blog post; all figures come from it, so read the official article for the full detail.

What each model is for

Haiku 5.5: fastest and cheapest, for classification, routing, bulk extraction and subagents that assist a larger model.

Sonnet 5.5: well-scoped everyday work such as a bug fix, a feature built to spec, chat, documents and spreadsheets.

Opus 5.5: the recommended starting point and Claude Code's default, the daily driver for complex coding and knowledge work.

Fable 5.2: the most capable generally available model, for major refactors and migrations, architecture planning, deep research, scientific work and long unattended runs.

Official prices at a glance

Haiku 5.5: $0.10 input and $0.50 output per million tokens, rising to $0.50 and $2.50 for prompts over 100K tokens.

Sonnet 5.5: $2 input, $10 output; $0.10 for cache reads.

Opus 5.5: $4 input, $20 output; $0.20 for cache reads.

Fable 5.2: $8 input, $40 output; $0.20 for cache reads, the same as Opus.

All four have a 1M-token context window and up to 128K output tokens; only Opus 5.5 and Fable 5.2 offer the faster, pricier fast mode.

Tune effort before switching models

Effort is a finer dial than changing models: higher effort means longer thinking, more tool calls and more self-checking, at more latency and cost.

From the article: on SWE-bench Pro, Opus 5.5 at medium scored about 2.5 points below high for about 70% of the cost, while xhigh added only about 1.4 points for 2.5 times the cost.

Past high, one more level can cost as much as the next model up, so test the bigger model at a lower effort instead.

Effort cannot add knowledge or skill a model lacks; for that you need a bigger model.

When results fall short, go in this order

First, let the model verify its own work with a test, a build or a script that calls the endpoint.

Second, raise effort by one level.

Third, move to the next model up, starting at a lower effort.

If outcomes can be checked automatically, run everything at low and re-run only failures at high. In the article this passed about 97% of tasks at about $0.17 each, versus 95.3% at $0.29 with everything at high.

Run subagents on smaller models

A subagent works in its own context and bills at its own model's rate. Give Haiku or Sonnet the work where mistakes are easy to spot, such as finding files, running tests and reading logs, and keep edits and judgment calls on the main model.

In Claude Code, set model: haiku in a subagent definition, or set a default with the CLAUDE_CODE_SUBAGENT_MODEL environment variable.

Note that the built-in Explore subagent now follows your session's model; define your own subagent with the same name pinned to haiku to keep exploration cheap.

For work bigger than one context window, a strong lead can split the job across many cheaper workers. In the article, an Opus 5.5 lead with up to 25 Sonnet 5.5 workers cost about a third less than Opus alone, with about the same score.

Things to check when switching models

The prompt cache belongs to the model that wrote it, so the first turn on a new model pays cache-write prices again. Run /compact in Claude Code first, or start a fresh session with a short plan.

Moving down a tier loses earlier reasoning, because a smaller model cannot read a larger model's thinking blocks.

Thinking cannot be turned off on Opus 5.5 and Fable 5.2, and Sonnet 5.5, Opus 5.5 and Fable 5.2 reject forced tool use, so send tool_choice as auto.

The most reliable method is to run your own real tasks on each candidate at two or three effort levels and compare cost and results per finished task. In Claude Code, switch with /model and check spend with /usage.

Original: Which Claude model should you build with? by Addy Osmani. Related product: Claude API. Related guide: Claude API credit vs token packs.