Model IDs and versioning

Two things about Claude model IDs surprise people, and both cost real time.

Every ID is pinned. None of them float.

claude-opus-5 looks like an evergreen pointer that will quietly follow the newest Opus release. It is not. Every Claude model ID is a pinned snapshot.

• Models with a date in the ID (claude-haiku-4-5-20251001) are fixed to that release, which is the obvious case.
• From the Claude 4.6 generation onward, IDs use a dateless format — claude-opus-5, claude-sonnet-5 — and these are also pinned snapshots, not pointers.

For models before the 4.6 generation, the “Claude API alias” column is a convenience pointer that resolves to a dated ID. From 4.6 onward, the alias column and the ID column are the same string.

The practical consequence: nothing auto-upgrades. Pinning was always the safe practice; now it is the only behaviour. A new Opus release will not reach your application until you change the string yourself — and the model you pinned will eventually retire.

The same model has a different name on every cloud

Model	Claude API	Amazon Bedrock	Google Cloud
Claude Opus 5	claude-opus-5	anthropic.claude-opus-5	claude-opus-5
Claude Sonnet 5	claude-sonnet-5	anthropic.claude-sonnet-5	claude-sonnet-5
Claude Haiku 4.5	claude-haiku-4-5-20251001	anthropic.claude-haiku-4-5-20251001-v1:0	claude-haiku-4-5@20251001

Note the third row: Bedrock appends -v1:0, Google Cloud uses @ before the date, and the Claude API uses a plain hyphen. Any code that constructs IDs by string manipulation across providers will eventually build one that does not exist.

Claude Platform on AWS is the exception that catches people out. Despite running on AWS, it uses Claude API-style IDs (claude-opus-4-6), not Bedrock-style ones — and follows Anthropic’s first-party deprecation schedule rather than Bedrock’s.

Query it instead of hard-coding it

The Models API returns every available model with max_input_tokens, max_tokens and a capabilities object. For anything that needs to adapt — a model picker, a token-budget check, a capability gate — that is more durable than a table copied into your source at some point last year.