Claude — AI Glossary

Reference details

Related terms

• Anthropic API
• Constitutional AI (CAI)
• Safety Training

How it works

1. The family is organised into size tiers that share a tokenizer, an API shape and behavioural conventions, so a request written against one tier runs against another by changing a model identifier.
2. Training combines supervised fine-tuning with preference optimisation and a constitutional method, in which the model critiques and revises its own drafts against written principles rather than against human labels for every comparison.
3. That written set of principles is the auditable artefact: the boundary the model enforces is documented rather than implicit in a preference dataset, which is the practical difference from a purely preference-trained model.
4. Dated snapshots name a specific trained artefact. The family and tier describe a lineage and a size; only the snapshot determines what will actually answer a request, which is what pinning in production means.
5. An extended thinking mode allocates tokens to reasoning before the response. Those tokens are generated and billed as output while not forming part of the visible answer, which changes the cost accounting rather than just the quality.
6. Access is available directly and through major cloud marketplaces, and the available tiers, snapshots and context limits differ by route, so the route is part of the configuration rather than a procurement detail.

Example: How extended thinking changes the token accounting

A request returns a 200-token visible answer. With extended thinking allocating 2,000 tokens of reasoning first, the billed output is 2,000 + 200 = 2,200 tokens — 2,200 / 200 = 11x the visible response. At 75 units per million output tokens, one such request costs 2,200 x 75 / 1,000,000 = 0.165 units against 200 x 75 / 1,000,000 = 0.015 without, and across 100,000 requests a month that is 16,500 against 1,500: a difference of 15,000 units. Whether that is worth paying is a per-task question with a measurable answer, and the arithmetic makes the shape of the decision clear. If extended thinking lifts task success from 80% to 92% on 100,000 requests, it converts 100,000 x 0.12 = 12,000 failures into successes for 15,000 units, so roughly 1.25 units per recovered task. That is cheap for a task whose failure costs a human intervention and expensive for one where a failure costs a retry. The mode belongs on the requests that are hard, which means routing rather than a global setting.

Commonly confused with

The family name is not a version and not a snapshot, and configurations that store only the family are how silent behaviour changes reach production. The family is the lineage, the tier is the size position within it, and the dated snapshot is the immutable artefact. Separately, the tier names are positions within this family only — they are not a cross-vendor scale, so comparing a tier name against another vendor's tier name is not a comparison of anything. For capability questions, the benchmark and provider pages hold the measured figures; this entry describes the mechanics that decide what you are running.

When to use it

Reach for it when:

• Planning a cost-and-capability ladder behind one integration, where tiers share the request format and the tokenizer.
• Pinning a dated snapshot for production, so prompt-sensitive behaviour does not shift under an unpinned identifier.
• Routing extended thinking to hard requests rather than enabling it globally, since the billed output multiple applies per request.
• Choosing an access route deliberately, as available tiers, snapshots and limits differ between direct and marketplace access.

Reach for something else when:

• As a capability claim on its own. Family membership predicts API shape, not performance on your task.
• Compared tier-for-tier against another vendor's naming, which is not a shared scale.
• With extended thinking on by default, where reasoning tokens are billed on every request including the easy ones.
• Configured at family or tier level in production, which leaves the actual artefact unpinned.

Continue your research

Use these connected ModelRefs sections to compare alternatives, inspect implementation paths, and review the evidence and governance boundaries relevant to Claude — AI Glossary.

• Browse models
• Review providers
• Inspect benchmarks
• Explore workflows
• Learn AI concepts
• Read decision guides
• Review trust and methodology
• Start a decision path