[model-inventory] Model alias inventory update - 2026-07-30 #49014

Summary

Analyzed live model inventories from OpenAI, Anthropic, Gemini, Copilot SDK, and Copilot (reflect) against the built-in alias mapping and pricing payload. The alias file (pkg/workflow/data/model_aliases.json) is already well-maintained and covers nearly all current model generations (sonnet-6x, gpt-5.6, gemini-3.x, etc.). One clear gap was found: grok-4.5 (xAI's Grok model, served through Copilot) has no alias coverage and is missing from models.json pricing data entirely, despite being actively served by Copilot with full billing metadata available.

• Providers queried: OpenAI, Anthropic, Gemini, Copilot SDK, Copilot (reflect)
• Total models found: 222 (OpenAI) + 10 (Anthropic) + 50 (Gemini) + 22 (Copilot SDK) + 37 (Copilot reflect)
• Proposed alias changes: 1
• Pricing gaps found: 1 (missing model)

Provider Model Counts

Raw API Fields Discovered

• Copilot SDK / reflect: billing.tokenPrices (inputPrice/outputPrice/cachePrice per-million-token, plus longContext tier pricing), capabilities.limits.max_context_window_tokens, capabilities.supports.vision/tool_calls/reasoning_effort, modelPickerCategory, modelPickerPriceCategory, policy.state.
• Anthropic: display_name, created_at, type — no direct pricing field (matches known family-based pricing tiers already in models.json).
• Gemini: input_token_limit, output_token_limit, supported_generation_methods, version — useful as a complexity/pricing-tier proxy, no direct cost field.
• OpenAI: owned_by, created — used for naming-convention cross-checks only.
• Docs pricing table (https://docs.github.com/en/copilot/reference/copilot-billing/models-and-pricing): confirms GPT family pricing tiers (Default vs Long context) matching models.json; does not list Anthropic/Gemini/xAI models (rendered separately or omitted from the public table).

models.json Pricing Analysis

Missing from models.json

Historical entries not currently returned

These model IDs remain in pkg/cli/data/models.json but are absent from all live inventories this run. Retained as historical records — no removal proposed:

claude-fable-5, claude-opus-4.5, claude-sonnet-4, gemini-2.5-pro, gemini-3-flash-preview, gpt-4.1, gpt-5.2, gpt-5.2-codex, gpt-5.4-nano (intentionally deprecated/excluded per policy), kimi-k2.7-code, raptor-mini

Inferred vs stored pricing discrepancies

None found. All GA models present in both models.json and current live inventories (gpt-5.4, gpt-5.5, gpt-5.3-codex, gpt-5-mini, claude-sonnet-4.5, claude-haiku-4.5, etc.) have pricing consistent with the Copilot SDK/reflect tokenPrices and the docs pricing table within rounding tolerance.

Proposed Alias Updates

1. What: Add a new grok alias family
Why: grok-4.5 is actively served by Copilot (confirmed in both Copilot SDK models and the live reflect endpoint model list) but has zero alias coverage today — it only matches the catch-all any alias, with no semantic or family-specific alias, and it's also absent from models.json pricing data.
Syntax:

{
  "grok": ["copilot/*grok*", "openai/*grok*"]
}

Also add grok-4.5 (and its github/grok-4.5 mirror) with the pricing shown above to pkg/cli/data/models.json and actions/setup/js/models.json (keeping the two mirrored files identical), using provider_type: "openai"-style entry consistent with other Copilot-routed non-Anthropic models (needs a maintainer decision on exact provider_type/wire_api since Grok is not currently modeled as its own provider in the schema — the two files today only have anthropic/openai/github-copilot provider buckets).

Notes

• pkg/cli/data/models.json and actions/setup/js/models.json remain byte-identical (verified via diff) — any pricing update must preserve this parity, e.g. via make refresh-models-json followed by manual addition of the Grok entry (models.dev does not yet list grok-4.5 under GitHub Copilot as of this run).
• gpt-4o-mini, gpt-4.1, gpt-4o, and gpt-5.4-nano continue to be treated as intentionally deprecated per standing policy and are excluded from the gap tables above even though they still appear in the live reflect model list.
• All other alias families (sonnet, sonnet-6x, haiku, opus, gpt-5* variants, gemini-* variants, kimi, mai-code, raptor-mini, task-oriented aliases like summarization, coding, reasoning, vision, image-generation) already have adequate wildcard coverage for all models observed in this run — no changes needed there.

network:
  allowed:
    - defaults
    - "accounts.google.com"
    - "android.clients.google.com"
    - "clients2.google.com"
    - "collector.githubapp.com"
    - "contentautofill.googleapis.com"
    - "safebrowsingohttpgateway.googleapis.com"
    - "www.google.com"

Generated by 📦 Daily Model Inventory Checker · aut00 · 50.6 AIC · ⊞ 12.5K · ◷

• expires on Aug 5, 2026, 4:18 PM UTC-08:00