AI Compass / Knowledge cards

Model choice

cost per task instead of price per token

Checked 11 October 2026 · Michael Engel · General · Costs

The token price tells you what a model charges per unit of processed text. What a finished task costs also depends on how many tokens the model needs for it. A current example with manufacturer figures for both quantities is Claude Opus 5.5, released on 22 Sept 2026.

What it is for

You avoid wrong decisions when switching models. The price list alone doesn’t tell you what your workflows end up costing.

Steps

1. Take 3–5 typical tasks from your everyday work as a fixed test set.
2. Run them with your current model and the new one, using the same settings.
3. For each task, measure the cost, the duration and whether the result is usable.
4. Separate by type of task, for example code, text or design. Often no single model wins everywhere.
5. Manufacturer figures and video comparisons are hints, not measurements. Don’t adopt them unchecked.

Example Opus 5.5 (manufacturer figures)

• List price: $4 per million input tokens and $20 per million output tokens. That is 20% less than Opus 5. Cache reads cost $0.20 per million tokens.
• Cost per typical task according to Anthropic: 40% less than Opus 5. That is the lower token price and the lower token consumption combined.
• Performance according to Anthropic: on most tasks at the level of Claude Fable 5.1.

The 20% and the 40% don’t contradict each other. One figure is the price, the other is consumption times price.

Links

• Decision models: for pure decisions, a large model is often not needed at all.
• Checkpoints and gates: each gate can use its own suitable model.
• Claude Code in the cloud: the place of execution also affects cost.
• Maintaining your Claude Code setup: in Claude Code you can set the model per type of task, for example with opusplan or a fixed model for subagents.
• Keeping context clean: the cost of a task also drops when fewer tokens are sent along.

Sources

• Anthropic: Claude Opus 5.5, 22 Sept 2026 (manufacturer’s prices and measurements)
• Jack Roberts: Opus 5.5 vs GPT-6 Astra: 7 Real Use Cases, 23 Sept 2026, chapter 12:29 (division of tasks; individual judgments, not verifiable; affiliate links in the description)

Status

Checked 2 Oct 2026, link added 9 Oct 2026 · Maturity new · Prices can change; check again before making a decision.

Want to try this with your team? That’s what the AI Lab is for. See the AI Lab