Claude Fable 5.2 or Opus 5.5? Anthropic's own advice: start with Opus 5.5 for most work, and move to Fable 5.2 for demanding reasoning, long-horizon agentic tasks, or when Opus 5.5 at higher effort still falls short (official models overview). Fable 5.2's API price is exactly twice Opus 5.5's.
This guide starts with a comparison table, then covers how to switch between the two models in Claude Code, how to have the main agent run Fable while subagents run Opus, and how to choose in Codex and Cursor. On November 12, 2026, 16:18–16:30 Beijing time (UTC+8), we ran the same bug-fix task on both models with Claude Code 2.1.322 and an AICodeWith API key. Timings and results are below.
The short answer
| Claude Fable 5.2 | Claude Opus 5.5 | |
|---|---|---|
| API model ID | claude-fable-5-2 | claude-opus-5-5 |
| Official positioning | Demanding reasoning and long-horizon agentic work | Long-running agentic coding and knowledge work; Anthropic recommends starting here for most workloads |
| Official API price (input / output, per million tokens) | $8 / $40 | $4 / $20 |
| Cache reads (per million tokens) | $0.20 | $0.20 |
| Context / max output | 1M / 128K | 1M / 128K |
| API default effort | high | medium |
| Official comparative latency | Moderate | Moderate |
| Minimum Claude Code version | v2.1.321 | v2.1.280 |
| Our test (same bug-fix task, Claude Code 2.1.322) | 2 runs, about 17.1 s and 18.9 s, 3 turns each, all tests passed | 3 runs, about 154.5 s, 327.7 s and 153.6 s, 4 to 8 turns, all tests passed |
Prices, specs and default effort come from Anthropic's pricing page and models overview. Minimum versions come from the Claude Code model configuration docs. Checked on November 12 at 16:20 Beijing time.
How to choose, in short:
- Everyday coding and well-scoped changes: start with Opus 5.5. It costs half as much per token.
- Unclear root causes, cross-module refactors, tasks that run for a long time: switch to Fable 5.2.
- Mix them: let a Fable main agent plan and orchestrate, and hand well-bounded subtasks such as writing tests or research to a subagent that runs Opus. You can also do the reverse: an Opus main agent that hands hard problems to a Fable subagent. Setup is below.
How Fable 5.2 and Opus 5.5 differ
The specs are almost identical: both have a 1M context window, 128K max output, always-on adaptive thinking and a June 2026 knowledge cutoff. The differences come down to three things:
- Positioning: Anthropic describes Fable 5.2 as being "for demanding reasoning and long-horizon agentic work" and Opus 5.5 as being "for long-running agentic coding and knowledge work". The Claude Code docs say Fable suits tasks larger than a single sitting, investigates before acting, and verifies its own work more often.
- Price: Fable 5.2 is $8 input and $40 output, Opus 5.5 is $4 and $20, exactly 2x. In fast mode, Fable 5.2 is $16 / $80 and Opus 5.5 is $8 / $40.
- Default effort: on the API, Fable 5.2 defaults to high and Opus 5.5 to medium. According to Anthropic's launch announcement, Fable 5.2 defaults to medium effort in Claude Code.
For Fable 5.2's pricing, benchmarks and 1M context on their own, see What is Claude Fable 5.2. For tuning Opus 5.5's effort, see Choosing an Opus 5.5 effort level. Still deciding between Sonnet and Opus? Read Sonnet 5.5 vs Opus 5.5 and Which model to use in Claude Code.
What developers are reporting
These are third-party claims, all from X story summaries and user reports:
- Agent leaderboard: according to an X story summary, Fable 5.2 took first place on Arena.ai's Agent Arena, leading in code, chat and work tasks. The same summary notes that some developers still prefer the lower-priced Opus 5.5 for routine work.
- Mixed routing: according to the X story "Developers Compare Claude Fable 5.2 and Opus 5.5 in Real Tests", developers broadly report that Fable 5.2 is stronger at agentic coding and long tasks but costs about twice as much, and many use Opus day to day while saving Fable for orchestration and hard problems.
- Subscription limits: according to user reports summarized on X, Fable burns through weekly limits on Claude subscriptions faster, so some people use Fable only for planning and give the heavy lifting to Opus.
- Reports of Opus slowing down: according to user reports summarized on X, some developers say that since November 3, Opus 5.5 in Claude Code makes more tool calls and takes longer per task. Anthropic has not commented.
These reports point the same way as our test below, but every sample here is small. Trust tests on your own project.
Our test: the same bug-fix task on both models
Setup: Linux x64, Claude Code 2.1.322, a global ~/.claude/settings.json pointing at an AICodeWith API key, non-interactive claude -p, --permission-mode bypassPermissions (only in a throwaway test directory; don't do this in a real repo).
Test project: a roughly 40-line Python inventory module. The README lists 3 rules, and the code breaks all 3 (adding an existing SKU overwrites the quantity instead of adding to it, a quantity that drops to 0 isn't removed, and the low-stock check uses < instead of <=). The 3 existing unittest tests all pass and catch none of this.
Prompt (identical for both models; we ran it in Chinese, translated here):
The spec is in README.md. inventory/stock.py doesn't match the spec in some places. Find and fix every mismatch, add one unittest test for each, then run python3 -m unittest to confirm everything passes. Finish with a one-line summary of what you changed.
Commands:
claude -p "<prompt above>" --model claude-fable-5-2 --output-format json
claude -p "<prompt above>" --model claude-opus-5-5 --output-format json
| Run | Model | Time (duration_ms) | Turns | Result |
|---|---|---|---|---|
| Round 1 (both models at once) | Fable 5.2 | about 17.1 s | 3 | Found all 3, fixed them, added 3 tests, Ran 6 tests ... OK |
| Round 1 (both models at once) | Opus 5.5 | about 154.5 s | 4 | Found all 3, fixed them, added 3 tests, all 6 tests passed |
| Round 2 (one after the other) | Opus 5.5 | about 327.7 s | 8 | Same, all 6 tests passed |
| Round 2 (one after the other) | Fable 5.2 | about 18.9 s | 3 | Same, Ran 6 tests ... OK |
| Round 3 | Opus 5.5 | about 153.6 s | 4 | Same, all 6 tests passed |
After every run we reran python3 -m unittest in the project directory ourselves. Every result was OK.
What we noticed:
- Both models got it right: they found all 3 mismatches, fixed them the same way, and put the tests in the existing test file. Opus wrote longer summaries that explain why each original line was wrong (for example, pointing out that
qty < 0can never be true). - Opus was clearly slower this time: we also sent two minimal "reply OK" requests to each model with curl. Opus 5.5 took about 1.2 s and Fable 5.2 about 2.9 s, so the slowdown came from Claude Code's multi-turn, long-context requests, not from the model being slow to respond at all. This only reflects this time window on this route and may differ at another time. Compare it with the "Opus slowing down" reports above.
- A small task doesn't show a capability gap: both models solved this one on the first try. Anthropic says Fable's advantage shows on harder, longer tasks, so compare on your own repo for that kind of work.
How to switch between Fable and Opus in Claude Code
1. Check your version: Fable 5.2 needs Claude Code v2.1.321 or later, and Opus 5.5 needs v2.1.280 or later. Run claude --version to check and claude update to upgrade.
2. Connect the API: put the following in your global ~/.claude/settings.json (%USERPROFILE%\.claude\settings.json on Windows). If the file already exists, merge these fields instead of overwriting it, and don't put them in project settings that get committed to the repo. The full steps are in Claude Code third-party API setup.
{
"model": "claude-opus-5-5",
"env": {
"ANTHROPIC_BASE_URL": "https://api.aicodewith.ai",
"ANTHROPIC_AUTH_TOKEN": "YOUR_AICODEWITH_API_KEY",
"ANTHROPIC_API_KEY": ""
}
}
Here "model" is set to Opus 5.5 as the everyday default. When we changed it to claude-fable-5-2 in our test, launching without --model used Fable 5.2 with a 1M context window and 128K max output.
3. Switch inside a session:
| What you want | How |
|---|---|
| Switch the current session to Fable 5.2 | /model fable or /model claude-fable-5-2 |
| Switch back to Opus 5.5 | /model opus or /model claude-opus-5-5 |
| Switch for this session only, keep the default | Type /model to open the picker, then press s on the model's row |
| Choose at launch | claude --model fable |
| Set a fixed default | The "model" field in settings.json |
According to the official docs, typing /model <name> or pressing Enter in the picker writes your choice to the model field in your user settings, so it becomes the default for new sessions. In our test through AICodeWith, --model fable and --model opus resolved to claude-fable-5-2 and claude-opus-5-5. To pin a version so an alias can't move to a newer model later, use the full ID.
Mixing them: Fable as the main agent, Opus for subagents
A custom subagent in Claude Code is a Markdown file with YAML frontmatter, stored in ~/.claude/agents/ (available in every project) or in a project's .claude/agents/. The model field in the frontmatter sets which model the subagent uses. It accepts the aliases sonnet, opus, haiku and fable, a full model ID, or inherit (same as the main conversation).
The subagent from our test, in .claude/agents/test-writer.md (translated from Chinese):
---
name: test-writer
description: Adds unittest tests for code that has already been fixed and runs them. Call it after the main agent finishes the bug fix.
tools: Read, Edit, Write, Bash
model: claude-opus-5-5
---
You only handle tests: read the spec in README.md and the changes, add one unittest test for each fix, run python3 -m unittest, and report how many passed. Don't change the business code under inventory/.
Then start the main session on Fable 5.2, let it fix the bugs itself and hand the tests to the subagent:
claude -p "Fix the places where inventory/stock.py doesn't match the spec yourself (don't write tests), then call the test-writer subagent to add a test for each fix and run python3 -m unittest." --model claude-fable-5-2 --output-format json
Result: about 74.0 s in total, 5 turns in the main session. The modelUsage field in the JSON output listed both claude-fable-5-2 and claude-opus-5-5, so both models were really called. Fable fixed all 3 issues, the Opus subagent added 6 tests, and all 9 tests passed.
Things to watch when you set this up (from the official subagents docs):
- Model resolution order: the
modelparameter passed at invocation → themodelfield in the subagent file → theCLAUDE_CODE_SUBAGENT_MODELenvironment variable → the main conversation's model. The invocation parameter only accepts aliases (such asopusorfable), not full IDs. - No
modelmeans it follows the main session: if you switch to Fable with/model, subagents without a model run on Fable too. To keep a subagent on Opus, setmodelin its file. - An alias can land on the main session's model: if the main session is already on an Opus model, a subagent set to
opususes that exact main-session model. A full ID such asclaude-opus-5-5is the safest choice. - Misspelled fields fail silently: Claude Code ignores fields it doesn't recognize. If a subagent doesn't switch models as expected, check the debug log with
claude --debug.
For more subagent model examples, see Configuring a Haiku 5.5 subagent in Claude Code and Background subagent model settings.
How to choose in Codex and Cursor
Codex: Codex uses OpenAI's GPT models, so Fable and Opus aren't in its model list. The same "everyday vs hard problems" split in Codex means switching between GPT models: codex -m <model> at launch, /model inside a session. To connect through AICodeWith, write your global ~/.codex/config.toml like this:
model_provider = "aicodewith"
model = "gpt-6.1-sol"
[model_providers.aicodewith]
name = "aicodewith"
base_url = "https://api.aicodewith.ai/chatgpt/v1"
wire_api = "responses"
experimental_bearer_token = "YOUR_AICODEWITH_API_KEY"
For how GPT compares with Opus, see GPT-6.1 Sol vs Claude Opus 5.5.
Cursor: pick the model from the dropdown in the Agent panel. According to Cursor's docs, Fable 5.2's input and output rates are about 2x Opus 5.5's. Accounts with Privacy Mode on and Enterprise customers need an admin to approve Anthropic's data retention terms in the dashboard first. When a request trips one of Fable's security guardrails, Cursor routes it to Claude Opus automatically.
Common errors and fixes
| Error / symptom | Cause | Fix |
|---|---|---|
| API Error: 400 {"error":"模型 claude-fable-5.2 不存在或未上线 ..."} (model does not exist or isn't live) | Wrong model ID. We deliberately wrote 5.2 instead of 5-2 and AICodeWith returned HTTP 400 | Use hyphens: claude-fable-5-2, claude-opus-5-5. Call GET https://api.aicodewith.ai/v1/models to list the models your key can use |
| [claude-code:unrecognized_model] at startup, and modelUsage shows a 200000 context window and 32000 max output | Claude Code older than v2.1.321 doesn't recognize Fable 5.2 (we hit this on 2.1.319) | Run claude update to get v2.1.321 or later; modelUsage should then show 1000000 / 128000 |
| There's an issue with the selected model | According to the official docs, the first request shows this when the model name in settings.json, --model or ANTHROPIC_MODEL is misspelled | Use the suggested name from the warning, or the full IDs from the table above |
| A subagent doesn't use the model you set | A misspelled frontmatter field is ignored, or the model parameter passed at invocation takes priority | Check the spelling of model, read the log with claude --debug, and use a full ID |
| 401 Unauthorized | Wrong or expired key, or wrong request headers | Work through the 401 troubleshooting guide |
| 429, or hitting usage limits quickly | Too much concurrency or rate limiting; Fable's higher per-token price uses up quota faster on the same work | Lower concurrency, wait and retry, and move everyday tasks to Opus |
The full 401 walkthrough is in How to fix Codex 401 Unauthorized, and 429 is covered in How to fix Codex 429 Too Many Requests. The same approach works for Claude Code.
FAQ
Is Fable 5.2 always better than Opus 5.5?
Anthropic positions Fable 5.2 for harder, longer tasks, and its benchmark scores are higher. But in our small task, both models got it right on the first try. For well-scoped everyday work, Opus 5.5 is enough at half the per-token price.
Which model does Claude Code use by default?
According to the official docs, no Fable model is the default on any plan or provider; you have to pick it. With a third-party API, set it explicitly with "model" in settings.json.
Can Claude Code switch between the two models automatically?
There's no switch that picks a model by task difficulty. The common approach is the subagent split in this guide: one model for the main session, and a subagent file that pins the other.
Can I use Fable 5.2 through AICodeWith?
Yes. On November 12 we queried /v1/models with an AICodeWith key and the list included claude-fable-5-2 and claude-opus-5-5. All the tests above ran through AICodeWith.



