include all prime inference models

[kevinjosethomas]

• the model picker only showed a hand-maintained subset of prime inference models, so users couldn't pick anything outside the whitelist.
• the catalog api returns just id and pricing, so we now borrow context window, max tokens, and vision from the openrouter metadata we already fetch, keep the curated table as overrides for prime's enforced limits, and fall back to safe defaults otherwise.
• the full catalog is now registered (95 models, up from 28) with raw and duplicate variants skipped.

---

Note

Medium Risk
Wrong context/max-token metadata can mislead the context bar or cause request failures, though overrides and conservative defaults reduce that risk; the generated model surface area grows substantially for Prime Inference.

Overview
Prime Inference moves from a whitelist of ~28 models to registering every catalog entry the API returns (raw/duplicate IDs filtered out), so the model picker can surface models that were previously omitted.

generate-models.ts now builds an OpenRouter metadata index and passes it into fetchPrimeInferenceModels. For each catalog row, context window, max tokens, and vision resolve in order: curated overrides (e.g. Anthropic capped at 200k without the 1M beta) → OpenRouter → catalog fields → conservative defaults (128k / 8192). Reasoning still uses the existing id/heuristic rather than OpenRouter’s flag.

Duplicate/noisy catalog IDs are skipped via isPrimeInferenceRawVariant (-BF16, zai-org/, uppercase vendor prefixes). Tests assert a larger catalog, OpenRouter-sourced metadata for non-curated models, and that filtered variants stay out.

models.generated.ts is regenerated: a large prime-inference section plus routine upstream tweaks (Bedrock EU costs, Fireworks/OpenRouter/Vercel gateway/opencode-go entries, etc.).

^{Reviewed by Cursor Bugbot for commit c151174. Bugbot is set up for automated code reviews on this repo. Configure here.}

Note

Include all Prime Inference catalog models instead of a curated whitelist

• Replaces the whitelist-only approach in fetchPrimeInferenceModels with a full catalog fetch, filtering only raw variants (e.g. -bf16 suffix) and duplicate vendor-prefixed aliases via isPrimeInferenceRawVariant.
• Adds buildOpenRouterMetadataIndex to cross-reference OpenRouter metadata for context window, max tokens, and vision flags, with curated overrides taking precedence over OpenRouter, which takes precedence over catalog fields and then hardcoded defaults (128k / 8192).
• Updates models.generated.ts with many new Prime Inference entries reflecting the full catalog.
• Behavioral Change: Previously only explicitly curated models were registered; now all non-duplicate catalog entries are included.

^{Macroscope summarized c151174.}

[linear]

ENG-4259 Include all Prime Inference models

Don't whitelist models, support all Prime Inference models

Review in Linear

[cursor]

Cursor Bugbot has reviewed your changes and found 1 potential issue.

^{❌ Bugbot Autofix is OFF. To automatically fix reported issues with cloud agents, enable autofix in the Cursor dashboard.}

^{Reviewed by Cursor Bugbot for commit c151174. Configure here.}

[cursor]

Sonnet 4 wrong context window

Medium Severity

Registering the full Prime catalog adds anthropic/claude-sonnet-4, but it has no entry in PRIME_INFERENCE_MODEL_METADATA, so createPrimeInferenceModel takes contextWindow from OpenRouter (1M). Other proxied Anthropic Sonnet/Opus entries are capped at 200k; the PR notes over-declaring the window breaks context budgeting.

Additional Locations (1)

• packages/ai/src/models.generated.ts#L13406-L13408

 

^{Reviewed by Cursor Bugbot for commit c151174. Configure here.}