Fable 5.2 costs 80× as much as Haiku 5.5 for input tokens and 80× as much for output tokens. Here is what that difference looks like on real workloads.
Prices per million tokens
| Item | Fable 5.2 | Haiku 5.5 |
|---|---|---|
| Input | $8 | $0.1 |
| Output | $40 | $0.5 |
| Cache write, 5 minutes | $10 | $0.125 |
| Cache write, 1 hour | $16 | $0.2 |
| Cache read | $0.2 | $0.01 |
| Batch API input | $4 | $0.05 |
| Batch API output | $20 | $0.25 |
Claude Haiku 5.5 has a higher price when a prompt is over 100,000 tokens: $0.5 input and $2.5 output per million tokens for the whole request.
Model ids: claude-fable-5-2 and claude-haiku-5-5.
The gap on real workloads
| Workload | Fable 5.2 | Haiku 5.5 | Difference |
|---|---|---|---|
| Chatbot, per message | $0.022 | $0.00030 | $0.021 |
| Support bot, per ticket | $0.186 | $0.0024 | $0.184 |
| Coding agent, per task | $2.40 | $0.036 | $2.36 |
| Document Q&A (RAG), per question | $0.068 | $0.00086 | $0.067 |
| Document summarization, per document | $0.280 | $0.0035 | $0.277 |
| Classification & extraction, per item | $0.0092 | $0.00013 | $0.0091 |
How to decide
- Measure, do not guess: run 50 to 200 of your real requests through both models and score the answers. Our guide to evals for Claude agents shows a small setup.
- Route by request: send routine requests to Haiku 5.5 and hand only the hard ones to Fable 5.2. Even a 70/30 split cuts most of the gap.
- Cache first: prompt caching often saves more than a model switch on agent and chat workloads.
Questions
Is Haiku 5.5 cheaper than Fable 5.2?
Yes. Haiku 5.5 costs $0.1 input / $0.5 output per million tokens, against $8 / $40 for Fable 5.2: 80× cheaper on input and 80× cheaper on output.
How much more does Fable 5.2 cost than Haiku 5.5 per month?
For a chatbot with 1,000 active users (300 messages each, with caching), about $6,449 more per month on Fable 5.2. The worked examples on this page show the gap for other workloads.
Should I use Fable 5.2 or Haiku 5.5?
Price is only half the answer. Run the same set of real tasks through both (a small eval of 50 to 200 examples is enough) and use Haiku 5.5 wherever it passes. Many apps route most requests to the cheaper model and send only hard ones to Fable 5.2.