Model cost is rarely what decides whether an AI system pays off. Rework, review time, and integration do. But the model bill scales with volume, and the difference between tiers is large enough to design around. This article uses Anthropic's published prices as of November 2026. Check the pricing page before you rely on any number here, because prices change.
The lineup
| Model | Positioning (Anthropic) | Input / output per million tokens |
|---|---|---|
| Claude Fable 5.1 | Demanding reasoning and long-horizon agentic work | $10 / $50 |
| Claude Opus 5.5 | Long-running agentic coding and knowledge work | $4 / $20 |
| Claude Sonnet 5.5 | The best combination of speed and intelligence | $2 / $10 |
| Claude Haiku 5.5 | High-volume, latency-sensitive work such as classification, extraction, and routing | From $0.10 / $0.50 |
All four have a 1M-token context window and up to 128K output tokens on the standard API. Anthropic's advice for choosing is to start with Opus 5.5 for most workloads, and move to Fable 5.1 when your evaluations at higher effort still fall short. The sensible follow-through is the reverse as well. Once an evaluation passes on a capable model, test cheaper tiers on the steps that allow it.
A worked example
Take a hypothetical workload: 1,000 documents, each using 20,000 input tokens and 1,000 output tokens. That is 20 million input tokens and 1 million output tokens. At list prices:
| Model | Input | Output | Total | With Batch (50% off) |
|---|---|---|---|---|
| Haiku 5.5 (from) | $2.00 | $0.50 | $2.50 | $1.25 |
| Sonnet 5.5 | $40.00 | $10.00 | $50.00 | $25.00 |
| Opus 5.5 | $80.00 | $20.00 | $100.00 | $50.00 |
| Fable 5.1 | $200.00 | $50.00 | $250.00 | $125.00 |
The spread between tiers is a factor of one hundred. That is why routing matters. If a small model can classify the document type and a mid-tier model can extract the fields, the expensive model is reserved for the cases that need it.
Three levers
Routing. Send each step to the cheapest model that passes its own evaluation. Classification and extraction belong at the Haiku tier. Reasoning over messy evidence belongs higher. Anthropic's pattern for this is the routing workflow described in the architecture article.
Batch. Anthropic lists batch requests at half price. The trade-off is asynchronous results, and Batch is not eligible for zero data retention or HIPAA arrangements, which the data handling article covers. It suits backlogs and overnight jobs on non-sensitive data.
Prompt caching. Cache reads cost a fraction of base input: 10 percent on Haiku 5.5, 5 percent on Opus 5.5 and Sonnet 5.5, and 2.5 percent on Fable 5.1. If 15,000 of the 20,000 input tokens in the example are the same instructions and reference text on every call, then on Sonnet 5.5 those cached tokens cost $1.50 instead of $30.00. The input bill falls from $40.00 to $11.50, and the total from $50.00 to $21.50. That figure excludes cache writes, which are priced separately, so treat it as an upper bound on the saving.
What does not show up in the table
- Effort and thinking. The current models use adaptive thinking steered by an effort setting, with different defaults per model. Higher effort changes quality, latency, and token use together. Measure it in your evaluation instead of guessing.
- Agents multiply tokens. Anthropic reports about four times the tokens of chat for agents and about fifteen times for multi-agent systems. A design that is cheap as a workflow can be expensive as a swarm.
- People. A review queue that routes 10 percent of cases to a person has a labor cost that usually dwarfs the model bill. Improving routing accuracy is often worth more than a cheaper model.
- Migration. Anthropic commits to retirement dates for each model. Pin the model ID, keep the evaluation, and budget a migration.
- Haiku tierclassify, route, extract
- Sonnet tierdefault reasoning
- Opus tierhard cases, long tasks
- Personwhat is left
The business case is therefore a short model: volume, cost per item at each tier, the share routed to people, and the cost of a person handling one. We build it before the system, from the evaluation results, so the number comes from measurement.
This is a reference design. It describes how we would build the system and cites Anthropic's documentation for platform facts. It is not a client case study, and the figures it uses are list prices or worked examples, not results.
Have a process like this?
Bring it to a 45-minute working session.