AI Development
Which model to use in Claude Code: the 3 rules from Anthropic’s official guide on Opus 5.5, Sonnet 5.5 and Fable 5.2
Anthropic says to start with Opus 5.5 at medium effort, use Sonnet 5.5 for well-defined work, and keep Fable 5.2, which costs twice as much, for big jobs.

Luca Di Domenico
· 5 min read
In short
Anthropic’s November 9, 2026 guide says to start with Opus 5.5 at medium effort, which is Claude Code’s default, switch to Sonnet 5.5 (half the price) for well-defined work, and keep Fable 5.2, which costs twice as much as Opus 5.5 per token, for very large jobs. If the result isn’t good enough, first give Claude a way to check its work, then raise effort by one level, and only then switch models. The Explore subagent now uses the session’s model instead of Haiku: a file in .claude/agents puts it back on Haiku, which I only recommend on the €20 plan.
On November 9, 2026, Anthropic published its official guide on which model to use in Claude Code. The short answer: start with Opus 5.5 at medium effort, drop to Sonnet 5.5 for well-defined work, and keep Fable 5.2 for very large jobs. If you run the most powerful model on every task, you burn through your plan’s limits for no reason.
The guide is called Which Claude model should you build with? Fable, Opus, Sonnet, or Haiku, written by Addy Osmani on Claude’s developer blog. Here I sum up the three rules that matter if you use Claude Code every day, plus one tip of my own.
How much does each Claude model cost?
API list prices per million tokens, according to the guide:
- Fable 5.2: $8 input, $40 output
- Opus 5.5: $4 input, $20 output
- Sonnet 5.5: $2 input, $10 output
- Haiku 5.5: $0.10 input, $0.50 output (for prompts up to 100,000 tokens)
The takeaway is simple: Fable 5.2 costs twice as much as Opus 5.5 per token, and Sonnet 5.5 costs half as much as Opus 5.5. If you use Claude Code on a subscription you don’t pay per token, but the same logic applies. Anthropic’s support page on Fable models says that on Max and Ultra plans, Fable models use up weekly limits faster than other Claude models.
Why can the Explore subagent eat your limits without you noticing?
When Claude Code needs to read your project, it hands the job to a subagent called Explore. Explore only gets read tools: it searches and opens files, but never edits them.
Explore used to always run on Haiku, the cheapest model. Since Claude Code version 2.1.198, Explore uses your session’s model. If you work with Opus 5.5, Explore also reads your project with Opus 5.5. If you work with Fable on a subscription, the subagents documentation says Explore runs on the Opus model. This is the trap few people know about: even simple file reading costs as much as the big model.
Rule 1: which model should you start with in Claude Code?
The guide’s first rule is to start with Opus 5.5 at medium effort. That’s already Claude Code’s default, so if you’ve never touched /model, you’re following it already.
For well-defined work, switch to Sonnet 5.5, which costs half as much as Opus 5.5. By well-defined I mean things like fixing a bug or adding a feature you’ve described very clearly. Keep Fable 5.2 for very large jobs, where the model has to reason for a long time across many parts of the project.
Rule 2: what should you do when Claude Code’s result isn’t good enough?
The second rule is not to switch models right away. The guide lists three steps, in this order:
- Give Claude a way to check its own work. For example a test, a build, or the ability to open the page and confirm it works.
- Raise effort by just one level.
- Only then move to a bigger model, starting at a lower effort level than you used on the smaller one.
Anthropic measured this approach on benchmarks. On SWE-bench Pro, Opus 5.5 at medium scored about 2.5 points lower than at high, for about 70% of the cost. At xhigh it scored about 1.4 points higher, but at 2.5 times the cost of high. Opus 5.5 at low, with only the failed tasks re-run at high, passed about 97% of tasks at about $0.17 each, against 95.3% at $0.29 with every task at high. Raising effort across the board costs more and delivers less than a setup that retries only where needed.
Rule 3: how do you put Explore back on Haiku?
The third rule moves Explore back to the cheap model. The guide explains that a subagent you create with the same name replaces Claude Code’s built-in one. Just ask Claude Code to create a file in .claude/agents that defines a subagent called Explore, running on Haiku and reading files without editing them. For example .claude/agents/Explore.md:
---
name: Explore
description: Searches and reads project files to answer questions about the code, without editing them.
tools: Read, Grep, Glob
model: haiku
---
Find and read the files needed to answer the question. Report the paths and the relevant lines. Don't edit any files.This part is my own addition: only put Explore back on Haiku if you’re on the €20 plan. On that plan the limits run out quickly, and saving on file reading makes a real difference. On the €100 or €200 plans, my advice is to keep using the session’s model, which reads your project more accurately.
I wrote about what changes with Opus 5.5’s lower price, if you want to understand why it’s now the model to start from.
Where to start tomorrow morning
The three rules fit in a few lines. Keep Opus 5.5 at medium as your starting point. Drop to Sonnet 5.5 when the task is clear. Before moving up to Fable 5.2, give Claude a way to check itself and raise effort by one level. If you run out of Claude Code limits after a single day, these are the first three settings to check.
Sources
- Which Claude model should you build with? Fable, Opus, Sonnet, or Haiku — Addy Osmani’s post of November 9, 2026: start with Opus 5.5 at medium; prices per million tokens Fable 5.2 $8/$40, Opus 5.5 $4/$20, Sonnet 5.5 $2/$10, Haiku 5.5 $0.10/$0.50 for prompts up to 100,000 tokens; order verify, effort, model; SWE-bench Pro (medium about 2.5 points lower at about 70% of high’s cost, xhigh +1.4 points at 2.5x the cost, low with re-runs at high about 97% at $0.17 vs 95.3% at $0.29); Explore on the session model and an Explore subagent with model: haiku
- Create custom subagents - Claude Code Docs — Docs read on November 12, 2026: Explore uses the main conversation’s model and only read tools; with Fable on a subscription it runs on the Opus model; a user or project subagent named Explore replaces the built-in one
- Claude Code CHANGELOG — Version 2.1.198: the built-in Explore agent inherits the main session’s model (capped at Opus) instead of running on Haiku
- Model configuration - Claude Code Docs — Docs read on November 12, 2026: for Pro, Max, Ultra, Team, Enterprise and API accounts the default model is Opus 5.5, which starts at medium effort
- Claude Fable models on your plan — Anthropic support page read on November 12, 2026, also covering Fable 5.2: on Max and Ultra plans, Fable models use weekly limits faster than other Claude models
Want to know what this means for your product?
If you're building software with AI and want a second opinion on how you're working, tell me about your project.
Tell me about your projectFollow the blog
One email for every new article
I only write when an article is out: title, summary and link. You can unsubscribe with one click from any email.
RSS feed
An RSS feed is the up-to-date list of articles: add it to a reader app like Feedly or Inoreader and new articles show up there, no email needed.