Claude Fable vs Opus vs Sonnet vs Haiku: API cost per task
Claude comes in four tiers: Fable at the top, then Opus, Sonnet, and Haiku. Each step down costs less per token and does less per call. Per-million-token prices are hard to compare in your head, so this post prices the same three tasks on each tier.
The prices
Official prices per million tokens, as listed on the AIKeys models page in October 2026:
| Model | Model id | Input | Output |
|---|---|---|---|
| Claude Fable 5.1 | cc/claude-fable-5-1 |
$10.00 | $50.00 |
| Claude Opus 5.5 | cc/claude-opus-5-5 |
$4.00 | $20.00 |
| Claude Sonnet 5.5 | cc/claude-sonnet-5-5 |
$2.00 | $10.00 |
| Claude 4.5 Haiku | cc/claude-haiku-4-5-20251001 |
$1.00 | $5.00 |
Every tier charges five times as much for output as for input. That has a useful consequence: the price gap between two tiers is the same whatever the mix of input and output. Fable costs 10 times Haiku, Opus 4 times, and Sonnet 2 times, on every task.
Opus 5.5 is also cheaper than the older Opus 4.7, which lists at $5 and $25. Moving to the newer model can lower the bill.
Three task shapes
- Classify: label a support ticket. About 500 input tokens, 20 output tokens.
- Answer: a chat reply with some context. About 2,000 input tokens, 500 output tokens.
- Agent step: one turn of a coding agent, with files and history in context. About 30,000 input tokens, 2,000 output tokens.
Cost per 1,000 tasks
| Task | Fable 5.1 | Opus 5.5 | Sonnet 5.5 | Haiku 4.5 |
|---|---|---|---|---|
| Classify (500 in, 20 out) | $6.00 | $2.40 | $1.20 | $0.60 |
| Answer (2,000 in, 500 out) | $45.00 | $18.00 | $9.00 | $4.50 |
| Agent step (30,000 in, 2,000 out) | $400.00 | $160.00 | $80.00 | $40.00 |
The working for one Opus 5.5 answer: 2,000 × $4 / 1,000,000 + 500 × $20 / 1,000,000 = $0.008 + $0.010 = $0.018. Times 1,000 is $18.
These are list prices. On AIKeys you pay the cheapest live ask, which is never more than half of these figures. The pricing explainer shows how to scale them to a month.
The agent column is the one to watch
Agents resend their whole context every turn, so input tokens dominate. A 50-turn coding session at 30,000 input tokens a turn is 1.5 million input tokens before any output. On Opus 5.5 that session costs about $8 at list price; on Fable 5.1, about $20. Trimming context, or caching it where the API supports caching, saves more than any other change.
Picking a tier
Start one tier lower than you think you need, and move up only when the cheaper model fails on your own examples.
- Haiku for high-volume, narrow work: classification, extraction, routing, short rewrites.
- Sonnet as the default for chat, writing, and most coding.
- Opus for long, multi-step reasoning and agent work where a wrong step is expensive.
- Fable for the hardest problems, where one better answer is worth ten times the tokens.
With one API key, a tier change is a different model string, so you can route each request to the tier it needs. The one-key setup guide has the code, and the models page lists every Claude id with its current lowest ask.