Claude Fable 5.2 is Anthropic’s most capable public model, and it costs twice as much per token as Claude Opus 5.5. Most of what you pay is decided by a few settings: which model you pick, the effort level, whether fast mode is on, and what your subagents run on. This guide walks through each one for Claude Code and the API, with commands you can copy.
Prices and settings are from Anthropic’s docs as of 15 November 2026. For what the model is and what changed in this release, read our Claude Fable 5.2 explainer first.
What you need
- Claude Code v2.1.321 or later (check with
claude --version, upgrade withclaude update), or an Anthropic API key inANTHROPIC_API_KEY. - A paid Claude plan or API billing. Fable models are not on the Free plan.
The price ladder at a glance
These are Anthropic’s list prices per million tokens from the pricing page.
| Option | Input | Output |
|---|---|---|
| Opus 5.5, standard | $4 | $20 |
| Opus 5.5, fast mode | $8 | $40 |
| Fable 5.2, standard | $8 | $40 |
| Fable 5.2, fast mode | $16 | $80 |
| Fable 5.2, Batch API | $4 | $20 |
Note that Opus 5.5 in fast mode has the same input and output price as Fable 5.2 at normal speed. Cached input is far cheaper than fresh input: a cache read on Fable 5.2 is $0.20 per million tokens against $8 uncached.
Step 1: Check that the task needs Fable
Anthropic’s advice in Choosing the right model is to start most work on Opus 5.5 and move to Fable 5.2 only when your tests at xhigh or max effort still fall short. Its cost guide agrees: for most agent workloads, start with Opus 5.5 at its default medium effort.
Fable suits agent sessions that run for hours, multistep research, root-cause debugging and architecture decisions. For everyday edits, summaries and chat, a cheaper model usually does the job.
Step 2: Know which effort level you are on
Effort controls how much the model thinks, how many tool calls it makes and how many tokens it uses. Fable 5.2 has five levels, and the default depends on where you use it:
- Claude API: the default is
high. - Claude Code: the default is
medium.
So the same model behaves, and bills, differently in the two places unless you set the level yourself. Anthropic’s prompting guide for Fable 5.2 describes the levels like this:
| Effort | Use it for |
|---|---|
low | Quick back-and-forth, well-specified tasks, high-volume work such as subagents |
medium | Interactive coding, chat, everyday knowledge work |
high | Hard coding, research, long unattended runs. Turns take noticeably longer |
xhigh | Your hardest long-running work, where you have measured a gain |
max | Measuring the model’s ceiling, or rare tasks where time and tokens don’t matter |
Anthropic says most of the capability gain is already there at high, and that max runs slower and costs more than xhigh, often for only a small gain.
Step 3: Set effort in Claude Code
Start a session on Fable at a chosen level:
claude --model fable --effort high
Or change it inside a session:
/model fable
/effort medium
Run /effort with no level to open a slider. Enter saves the level as your default for that model; s applies it to this session only. /effort auto clears your saved level. Levels are stored per model, so raising Fable to high does not change Opus. See the model configuration docs.
Step 4: Set effort on the API
Set output_config.effort on the request. This example runs Fable 5.2 at medium and streams the reply:
curl https://api.anthropic.com/v1/messages \
-H "x-api-key: $ANTHROPIC_API_KEY" \
-H "anthropic-version: 2026-10-21" \
-H "content-type: application/json" \
-d '{
"model": "claude-fable-5-2",
"max_tokens": 64000,
"stream": true,
"output_config": { "effort": "medium" },
"messages": [
{ "role": "user", "content": "Find the race condition in this handler and propose a fix: ..." }
]
}'
Expected output: a stream of server-sent events ending with a message_stop event.
Don’t try to save money with a small max_tokens. It is a hard limit on thinking plus the reply, the model can’t see it, and a cut-off turn is still billed. In Anthropic’s internal tests on repository tasks, a 16,384 cap ended 85% of Fable 5.2’s attempts and it solved 20.3% of tasks. At 64,000 it solved 77.4%, at a lower cost per solved task. Anthropic recommends 64,000 for agentic work.
If your work can wait, the Batch API halves the price, with results within 24 hours. Batch can’t be combined with fast mode.
Step 5: Use fast mode only where speed pays
Fast mode gives up to 2.5 times the output speed on the same model at double the token price. In Claude Code:
/fast
Press Space to toggle and Enter to confirm. A ↯ icon shows next to the prompt while it is on. On the API, add "speed": "fast" to the request and check usage.speed in the response.
Four rules keep the bill sane:
- Turn it on at the start of a session. The first time you enable fast mode in a conversation, the whole context is billed at the uncached fast-mode input price. Fast and standard speed keep separate caches, so each later switch re-reads context uncached.
- Remember it is extra on subscriptions. On Pro, Max, Ultra, Team and Enterprise plans, fast mode is billed to usage credits, even if you have plan usage left.
- Run
/fastwhile you are on Fable. Fast mode turns off when you switch to Fable 5.2 unless you turned it on there. - Try lower effort first. Lower effort is also faster, and it costs less instead of more.
Step 6: Stop subagents from running up the bill
Anthropic notes that Fable 5.2 delegates readily in agentic coding tools and that subagents can account for a large part of a session’s cost. By default a subagent runs on the same model as your session, so a Fable session means Fable subagents.
To pin one subagent to a cheaper model, add a model line to its definition file:
---
name: test-runner
description: Runs the test suite and reports failures
model: haiku
---
To put every subagent on one model, add this to the env block of a settings file:
{
"env": {
"CLAUDE_CODE_SUBAGENT_MODEL": "haiku",
"CLAUDE_CODE_SUBAGENT_MODEL_FORCE": "1"
}
}
Both options are in the subagents docs. Skills also keep context small because they load only when needed; see our guide to agent skills in Claude Code.
Step 7: Cap and watch your spend
Inside Claude Code, /usage shows the session’s token use and estimated cost. Run /clear between unrelated tasks, because every message re-sends the conversation so far.
For scripted runs, set a ceiling:
claude -p --model fable --max-budget-usd 5.00 "Audit src/ for unused exports and list them"
The cap works in print mode only and uses Claude Code’s own cost estimate, so leave headroom. On a subscription, -p runs never show the usage-credits consent prompt: a Fable request that bills to credits is billed without asking. API organizations can also set workspace spend limits in the Claude Console.
Know how your plan bills Fable. Per Anthropic’s help article, Max, Ultra and premium Team and Enterprise seats can spend up to 50% of their weekly limit on Fable, which uses that limit faster than other models. Pro and standard seats pay through usage credits. A launch offer that includes Fable on Pro and standard Team seats ends on 16 November 2026 at 11:59 pm PT; after that those seats use usage credits.
Common errors and fixes
- “Requires usage credits” on the Fable row in
/model: your plan bills Fable to usage credits. Run/usage-creditsto manage them. - “Fast mode requires usage credits”: turn usage credits on in Settings > Usage on claude.ai. On Team and Enterprise, an Owner must also enable fast mode.
- Fast mode switched off after
/model fable: expected. Run/fastagain while on Fable. - 429 errors with
"speed": "fast": fast mode has its own rate limit. Wait for theretry-aftertime, or resend withoutspeed, which misses the prompt cache. - Replies stop mid-task: raise
max_tokens. If the model stops and offers to continue, which is more common atlowandmedium, raise effort or add the “finish the whole task” text from Anthropic’s prompting guide. - Simple questions feel slow: at
highand above Fable thinks before almost every reply. Asking it to “answer directly” did not stop that in Anthropic’s testing, so lower the effort. - Everything fails, not just fast mode: see our guide on what to do when Claude is down.
Wrap-up
Treat Fable 5.2 as the model you escalate to. Start on Opus 5.5, move up when a task is hard and long, set effort on purpose, keep max_tokens generous, and put cheap models on side jobs. Save fast mode for when your own time costs more than the tokens.
More on Claude: see all our Claude and Anthropic guides, grouped by task.
Sources
Links last checked on 15 November 2026.





Write a response