# Built-in Catalog Summary — ModelCatalogData.swift

Generated 2026-07-30 from the per-provider research files in docs/research/. Audit companion for Sources/ZyquoCloud/Services/ModelCatalogData.swift — use this to detect drift between research and the shipped catalog. Total shipped: 195 chat models.

Global exclusion policy (applies to every provider): embeddings, audio (TTS/ASR/realtime/omni-audio), image/video generation, moderation, OCR, robotics, computer-use, deep-research agent products, and Responses-API-only models are never shipped.

Provider	Included	Excluded / trimmed	Why excluded
OpenAI	29	~100 live IDs	All `-pro` (gpt-5.5-pro, gpt-5.4-pro, gpt-5.2-pro, gpt-5-pro, o3-pro, o1-pro), all `-codex`, `deep-research*` (Responses API only); `ra-gpt-5.6-sol` (undocumented/experimental); `gpt-5-search-api`, `gpt-4o[-mini]-search-preview` (search-fee SKUs, unverified pricing); dated snapshots (alias kept); embeddings/audio/image/realtime/moderation/sora/davinci/babbage IDs
Anthropic	11	1+	`claude-mythos-5` (invite-only, Project Glasswing); dated aliases of 4.5-family kept as canonical dated IDs per research table
xAI	6	4	`grok-imagine-image[-quality]`, `grok-imagine-video[-1.5]` (image/video gen); retired models (grok-4, grok-4-fast, grok-3, grok-2-vision) absent from live `/models`; dated variants shipped via stable aliases (grok-4.20 etc.)
Mistral	10	~10	`voxtral-` (audio), `mistral-embed`, `codestral-embed`, `mistral-moderation-`, `mistral-ocr-`, `labs-leanstral-`, `mistral-vibe-cli-*` (product aliases); dated snapshots (kept `-latest` aliases only); `magistral-small-2509`/`mistral-small-2506` (deprecated 07-31, replaced by mistral-small-latest); `mistral-medium-2508/2505` (dated legacy snapshots)
Gemini	17	~25	TTS, image (incl. nano-banana), Imagen/Veo/Lyria, embeddings, Live/native-audio, robotics, computer-use, `aqa`, `antigravity-preview`, `deep-research-*` (agent products); trimmed previews: `gemini-3.1-pro-preview-customtools`, `gemini-3.1-flash-lite-preview`, `gemini-omni-flash-preview` (redundant/unverified preview channels)
Qwen	32	~119 of 151 live	Image/video (`qwen-image`, `wan`, `z-image`), TTS/ASR, omni/realtime/s2s, live-translate, MT (`qwen-mt-`), OCR, embeddings; all dated snapshots (aliases kept); `ccai-pro` (unidentified), `qwen3.6-max-preview`, `glm-5.2-fast-preview` (previews), `deepseek-v3.2`, `glm-5.1` (superseded third-party), `qwen-vl-max/plus` (legacy VL, superseded by qwen3-vl), `qwen3-235b-a22b` (streaming-only hybrid), small open weights (qwen3-32b/30b/14b/8b, qwen3.5-27b, qwen3.6-35b/27b, qwen2-7b) trimmed per curation cap
DeepSeek	2	2 retired	`deepseek-chat` and `deepseek-reasoner` retired 2026-07-24 (no longer resolve) — only the V4 pair exists
Kimi	12	0	All 12 live models are chat models; nothing excluded
Perplexity	4	1	`sonar-reasoning` (non-Pro) removed from the current docs enum; Agent API models out of scope (different endpoint)
Together	34	~144 of 178 chat/language entries	Dedicated-endpoint / 0-priced artifacts (GLM-4.7-fp4, GLM-5-FP4, Qwen 35B FP8 variants, MiniMax-M2, pearl-ai mirrors…); trimmed niche/duplicated: Kimi-K2.5-fp4, MiniMax-M2.7, GLM-4.5-Air, Llama-3.1-405B (probe ctx quirk), Llama-3.1-70B (superseded by 3.3), Llama-3.2-3B, Qwen2.5 family, Mistral-Small-24B-2501, gemma-3n, Nemotron-Nano-9B, cogito, LFM2.5, trinity-mini, Hermes/roleplay models, Llama-Guard (moderation)
DeepInfra	35	~139 of 174	Image/video/TTS/ASR entries (null context/pricing); trimmed duplicates/niche: claude-opus-4-7, claude-sonnet-4-6 (older proxied gens), DeepSeek-V3.2/V3.1-Terminus/V3-0324, GLM-5.1/4.7-Flash/4.6, Qwen3.6/3.5 mid-size variants, Qwen3-Max[-Thinking], Qwen3-Next-80B, Qwen3-VL-30B, phi-4, Mistral-Nemo, gemma-3-27b, Hy3, Step-3.7-Flash, MiMo, Seed-2.0, Inkling, Nemotron Super/Nano, Hermes/MythoMax/roleplay, Llama-Guard
Cerebras	3	0	Live catalog is exactly 3 public models; `zai-glm-4.7` shipped flagged legacy (discontinuation 2026-08-17)

# Flag conventions used

isRecommended (max 4 per provider): OpenAI gpt-5.6-sol/terra; Anthropic claude-opus-5/sonnet-5; xAI grok-4.5/grok-code-fast-1; Mistral medium/large/small-latest; Gemini 3.6-flash/3.5-flash-lite/3.1-pro-preview; Qwen qwen3.7-max/plus/flash; DeepSeek both; Kimi k3/k2.7-code; Perplexity sonar/sonar-pro; Together Kimi-K3/DeepSeek-V4-Pro/gpt-oss-120b; DeepInfra DeepSeek-V4-Pro/V4-Flash/GLM-5.2/gpt-oss-120b; Cerebras gpt-oss-120b.
isLegacy: OpenAI gpt-4.1/4o/4/3.5 families, gpt-4-turbo, o1, o3-mini; Anthropic 4.5-and-earlier dated models; Mistral magistral-medium/devstral/open-mistral-nemo; Gemini 2.0 family; Qwen qwen-turbo; Kimi moonshot-v1 family; Together Mixtral-8x7B; Cerebras zai-glm-4.7.

# Notable data-entry decisions (verify in Phase 7)

OpenAI reasoning models use ParameterSupport(temperature: false, topP: false, usesMaxCompletionTokens: true, reasoningEffort: true); *chat-latest models use .openAIDefault.
Anthropic 4.7+/5-generation models: sampling disabled, thinkingToggle: true; Fable 5 has thinkingToggle: false (thinking always on, cannot be disabled); 4.6-and-earlier keep temperature/topP.
Kimi: research documents reasoning_effort only on kimi-k3 (K2.5/K2.6 use the thinking object → thinkingToggle; K2.7-code thinking is always-on, no toggle/effort) — this deviates deliberately from the blanket "K-series reasoningEffort" guidance to match the documented API. K-series hides sampling sliders and prefers max_completion_tokens.
Qwen third-party pricing is unpublished on the intl docs → pricing: nil for most Qwen entries (3p figures kept only for the qwen3.7 tier); enable_thinking mapped to thinkingToggle.
Perplexity: all 4 models carry citations: true, no vision, no tools; context windows are tracker-sourced (unverified).
Cerebras context windows use the paid tier (131K); free tier caps at 65K.
DeepSeek pricing uses cache-miss input rates; both models get reasoningEffort + thinkingToggle.
Together/DeepInfra maxOutputTokens is nil throughout (listings expose no distinct output cap).