Part of #718.
Rule: a Bedrock model card's Programmatic Access table states the model id as <maker>.<model>, and its inference-profile ids are that id behind a geography (us., eu., apac., au., ca., jp., in., global., us-gov.). upstreamModelIdentity drops the geography, takes the maker before the first dot as the namespace, and strips Bedrock's own version (-v1:0, -v1, -1:0, then any trailing :N). Only version 1 is read as Bedrock's, so a maker's own -v4 or -v2 stays (cohere.embed-v4:0 -> cohere/embed-v4). Source: docs-derived, the model cards index URL, path Programmatic Access: Model ID (same shape as Azure).
Examples:
anthropic.claude-opus-4-5-20251101-v1:0 -> anthropic / claude-opus-4-5-20251101
global.anthropic.claude-opus-4-6-v1 -> anthropic / claude-opus-4-6
us.xai.grok-4.7 -> xai / grok-4.7 (grok claims xai)
us-gov.openai.gpt-oss-120b-1:0 -> openai / gpt-oss-120b
Counts against the 270 live rows (2026-11-21): 244 state an upstream, 126 of those have a target row today (anthropic 74 of 89, openai 41 of 53, xai 7 of 7, zai 4 of 4). Claude goes from 0/89 to 74/89.
Unlinked, and why:
amazon. ids (26 rows: Nova, Titan) state no upstream: Bedrock is their first-party host.
- Claude 3 Haiku, Sonnet 4 (20250514) and the Mythos rows: no Anthropic row exists.
moonshotai. (9 rows): no provider claims the moonshotai namespace, and both moonshot and moonshotai-cn could, so I left it out of this PR.
- Gemma, Llama, Qwen, Mistral, Cohere, NVIDIA, MiniMax, DeepSeek, Stability, TwelveLabs, Writer, Luma: Bedrock's model name differs from the maker row (
gemma-4-31b vs gemma-4-31b-it) or the maker has no row. Name matching is #719.