Claude Opus 5.5 vs Sonnet 5.5 vs Fable 5.1: Which to Use, and How to Split Them in Claude Code

Short answer

If you're not sure, use Opus 5.5 — Anthropic's own guidance is to start with Opus 5.5 for most workloads. Hand high-volume, latency-sensitive, not-too-hard work to Sonnet 5.5, which costs half as much. Move to Fable 5.1, at 2.5 times the Opus price, for problems Opus 5.5 still can't solve at a higher effort level, or for tasks that need to run autonomously for a long time.

A common Claude Code setup: Opus 5.5 as the main model, Sonnet 5.5 for subagents, and /model fable when you hit a hard problem. All three are available on TeamoRouter at roughly 30% of the list price (as of writing — the live numbers are on the pricing page).

Side by side

Prices are USD per million tokens from Claude's pricing page; cache write is the 5-minute cache rate. Positioning, speed and context come from Anthropic's models overview.

How far apart the prices are: 1 : 2 : 5

At list price, Sonnet 5.5, Opus 5.5 and Fable 5.1 are priced 1 : 2 : 5, for input and output alike. A typical Claude Code session on a mid-size project reads about 200k input tokens and writes about 20k (ignoring cache for now):

On TeamoRouter all three are roughly 30% of list, so the ratio stays about the same.

Two things shift that ratio:

• In long sessions most input is a cache read. Fable 5.1's cache read is $0.25, only 25% more than Opus 5.5's $0.20. The real gap between them is output ($50 vs $20) and the first cache write ($12.50 vs $5).
• Thinking is billed as output. The higher the effort level, the more the model thinks and the more output tokens you pay for. Anthropic notes that Opus 5.5 thinks more per turn than Opus 5 at the same effort, most of all at xhigh and max.

Which task goes where

Opus 5.5 is the default. Anthropic's guidance is to start here when unsure. Changes that span many files, getting to grips with an unfamiliar codebase, and agentic work where the model plans and then verifies its own steps all fit. Its default effort in Claude Code is medium; on a hard problem, try /effort high or /effort xhigh before switching models.

Sonnet 5.5 is for volume and speed. Anthropic rates it "fast", and it costs half as much as Opus. Writing tests, making changes from clear instructions, repetitive edits across a codebase, and conversational Q&A are cheaper here.

Fable 5.1 is for when Opus 5.5 at higher effort still falls short. Anthropic recommends it for demanding reasoning and long-horizon agentic work, or when your results on Opus 5.5 at higher effort aren't good enough. The Claude Code docs describe Fable as suited to tasks larger than a single sitting: it sustains long autonomous sessions, investigates before acting, and verifies its work more often than smaller models. The cost is that it's slower and more expensive.

Haiku 5.5 is for background chores and short requests. List price is $0.10 / $0.50, but a request over 100k tokens costs five times as much ($0.50 / $2.50), so it's a poor fit for large contexts. See the Haiku 5.5 guide.

We don't quote benchmark scores here; long-task behaviour is as Anthropic describes it. For your own project, the reliable test is to run the same task on two models and compare both the result and the bill.

Setting it up in Claude Code

Step 1: pin all four aliases to the current generation

Create a TeamoRouter API key in the console. Then pin all four Claude Code model aliases, so that opus, sonnet, haiku and fable in /model map to the models below. Add to ~/.zshrc (or ~/.bashrc if you use bash):

export ANTHROPIC_BASE_URL=https://api.teamorouter.com
export ANTHROPIC_AUTH_TOKEN="your TeamoRouter key"
export ANTHROPIC_MODEL=claude-opus-5-5
export ANTHROPIC_DEFAULT_OPUS_MODEL=claude-opus-5-5
export ANTHROPIC_DEFAULT_SONNET_MODEL=claude-sonnet-5-5
export ANTHROPIC_DEFAULT_HAIKU_MODEL=claude-haiku-5-5
export ANTHROPIC_DEFAULT_FABLE_MODEL=claude-fable-5-1

On Windows, in PowerShell, then close and reopen the terminal:

setx ANTHROPIC_BASE_URL "https://api.teamorouter.com"
setx ANTHROPIC_AUTH_TOKEN "your TeamoRouter key"
setx ANTHROPIC_MODEL "claude-opus-5-5"
setx ANTHROPIC_DEFAULT_OPUS_MODEL "claude-opus-5-5"
setx ANTHROPIC_DEFAULT_SONNET_MODEL "claude-sonnet-5-5"
setx ANTHROPIC_DEFAULT_HAIKU_MODEL "claude-haiku-5-5"
setx ANTHROPIC_DEFAULT_FABLE_MODEL "claude-fable-5-1"

Two easy mistakes: ANTHROPIC_BASE_URL must not end in /v1, and Claude Code must be recent enough — the docs require v2.1.280 or later for Opus 5.5, v2.1.284 for Sonnet 5.5, and recommend v2.1.293 for Haiku 5.5. Run claude update if you're behind.

Step 2: the main model

ANTHROPIC_MODEL=claude-opus-5-5 starts every session on Opus 5.5. Switch any time with /model sonnet or /model fable, and come back with /model opus.

Step 3: subagents

Claude Code hands some work to subagents that run in their own context — Explore, which searches the codebase, is the one you'll see most. Per the docs, a subagent uses the same model as the main conversation by default: if your main model is Opus 5.5, subagents are billed at Opus 5.5 too. Two ways to make them cheaper.

Set a default model for subagents:

export CLAUDE_CODE_SUBAGENT_MODEL=claude-sonnet-5-5

It's only a default: a model in a subagent's definition, or a model Claude passes when it starts one, wins. On its own, it also doesn't change the built-in Explore and Plan subagents, which keep following the main model.

Override the built-in Explore. Create .claude/agents/Explore.md in your project; a subagent with that name replaces the built-in one and uses the model you set:

---
name: Explore
description: Read-only search and reading of the codebase; never edits files
model: sonnet
tools: Read, Grep, Glob
---
Find the files and code relevant to the task quickly. Return file paths and the key snippets. Do not modify any files.

On a small repository you can set model to haiku and save more. On a large one, Explore often reads enough files in one request to cross 100k tokens and trigger Haiku 5.5's 5× rate, so Sonnet 5.5 is the safer choice.

To run every subagent, Explore and Plan included, on one model, also set CLAUDE_CODE_SUBAGENT_MODEL_FORCE=1 (v2.1.257 or later). The model field in subagent definitions is then ignored.

Optional: opusplan

/model opusplan uses Opus in plan mode and switches to Sonnet automatically when you leave plan mode to execute. With Step 1 in place, that means claude-opus-5-5 and claude-sonnet-5-5. Every switch in or out of plan mode is a model switch and rebuilds the cache — the next section explains why that matters.

Common setups

On the last row: when the main model is Fable, the docs state that Claude Code runs the built-in Explore on Opus instead, including when you connect to a gateway through ANTHROPIC_BASE_URL. To keep other subagents off Fable pricing, set CLAUDE_CODE_SUBAGENT_MODEL to claude-opus-5-5.

The hidden cost of switching mid-conversation

The Claude Code docs are explicit: each model has its own cache. After a /model switch, the next request reads the entire conversation history with no cache hits.

An example. Your conversation has grown to 200k tokens:

• Stay on Opus 5.5, and those 200k tokens are a cache read on the next turn: $0.04 at list price.
• Switch to Sonnet 5.5, and they have to be written to Sonnet's cache from scratch: $0.50 at list price.

The switch itself costs an extra $0.46. After that, at about 2k output tokens per turn, Sonnet 5.5 saves roughly $0.04 per turn over Opus 5.5 — about twelve turns to break even.

So don't flip back and forth in a long conversation. If you want the cheaper model, choose it at the start of a new session, or give the cheap work to a subagent: subagents have their own context and don't force the main conversation to re-read its history. This is also why Claude Code asks you to confirm a /model switch while the cache is still warm.

Using them on TeamoRouter

• All three models are available; IDs are in the table above, and each has a 1M-token context.
• Prices are roughly 30% of list as of writing; the live numbers are on the pricing page. Opus 5.5 pricing details and API examples are in Claude Opus 5.5 API: pricing and setup, and specs are on the Opus 5.5 model page.
• You pay only for tokens used — no subscription or monthly fee. Top up by card via Stripe, or with USDT — see how to pay with USDT.

FAQ

/model fable says the model isn't available. Check that ANTHROPIC_DEFAULT_FABLE_MODEL=claude-fable-5-1 is set, or type /model claude-fable-5-1 directly.

I set CLAUDE_CODE_SUBAGENT_MODEL but Explore still follows the main model. That's by design: on its own the variable doesn't change the built-in Explore and Plan. Create an Explore subagent as shown above, or also set CLAUDE_CODE_SUBAGENT_MODEL_FORCE=1.

Can I turn off thinking on Opus 5.5 to save money? No. Anthropic states that thinking is always on for Opus 5.5. To spend less, lower the effort, for example /effort low.

Do Sonnet 5.5 and Opus 5.5 have the same context window? Yes — 1M tokens each, with 128K max output. Choose by difficulty and budget, not by context.

Is opusplan cheaper than staying on Opus? Execution output is billed at Sonnet rates, so usually yes, but every switch in or out of plan mode rebuilds the cache. It suits tasks you plan once and then execute for a long time, not ones where you toggle constantly.

Next steps

• Claude Opus 5.5 API: pricing and setup — model ID, prices, Claude Code configuration.
• Use Opus 5.5 in Cursor, Cline and Cherry Studio — what to enter in third-party clients.
• Claude subscription or pay-as-you-go API — the monthly math.
• Create an API key — and connect Claude Code in five minutes.