Reference Links
Model Reference
Canonical September 2026 model IDs for every provider - the single source of truth for the model-ID integrity check.
Last verified 2026-09-24
This is the canonical list of model IDs Agent Surface uses in examples. Every model-ID string in the docs and in the template kits must appear here. scripts/check-docs-integrity.mjs parses this page's code-formatted IDs into an allowlist and flags any model-ID-shaped string elsewhere in the corpus that is not on it - including bare Anthropic role aliases (opus-4-7) and prose model names (Opus 4.8). Update this page first, then run the sweep.
Verified against provider documentation on 2026-09-24.
Anthropic
Anthropic also lists Mythos, a restricted tier above Fable 5.1 gated to an allowlist of enterprise accounts. No public API ID is published for it; do not reference it in general examples.
Retired 2026-06-15: Opus 4 and Sonnet 4. Retired 2026-08-05: Opus 4.1. Older pins from these eras - Opus 4.7, Opus 4.8, and Sonnet 4.6 - are no longer valid; the model-ID check flags them, including bare aliases and prose mentions. Superseded 2026-09-22: Opus 5 remains available in the API but is no longer the recommended default; Fable 5 remains available but is no longer the flagship pin. Replace default pins with claude-opus-5-5, flagship pins with claude-fable-5-1, and balanced pins with claude-sonnet-5.
OpenAI
| Model | API ID | Role | Notes |
|---|---|---|---|
| GPT-5.6 Sol | gpt-5.6-sol | Flagship | GA 2026-07-09. Deepest reasoning tier; the bare alias gpt-5.6 routes to Sol. |
| GPT-5.6 Terra | gpt-5.6-terra | Balanced | Mid tier - most capability per dollar for everyday agent work. |
| GPT-5.6 Luna | gpt-5.6-luna | Value tier | Fastest and cheapest; the default wherever a cheaper OpenAI model is wanted. |
Superseded 2026-07-09: GPT-5.5 and GPT-5.4 remain available but the GPT-5.6 tiers replace them in examples - the model-ID check flags remaining GPT-5.5/5.4 pins. The o-series reasoning models (o1, o3, o4 family) were deprecated in June 2026, with removal scheduled for December 2026. Do not use them in examples.
| Model | API ID | Role | Notes |
|---|---|---|---|
| Gemini 3.7 Flash | gemini-3.7-flash | Current | GA 2026-08. Default Google model for agent and eval examples. |
| Gemini 3.6 Flash | gemini-3.6-flash | Previous | Stable previous-generation Flash model; retain only for deliberate compatibility pins. |
| Gemini 3.5 Flash-Lite | gemini-3.5-flash-lite | Fast | GA 2026-07-21. Fastest 3.5-series model for high-volume, low-cost calls. |
| Gemini 3.1 Pro (preview) | gemini-3.1-pro-preview | Preview | Higher-capability preview; still preview as of 2026-09-24 - pin the full ID. |
Superseded 2026-08: Gemini 3.7 Flash replaces Gemini 3.6 Flash as the example default. Keep gemini-3.6-flash only where compatibility is deliberate. (Gemini 3.5 Pro is not listed because it has not reached GA.)
Embeddings
| Model | API ID | Provider | Notes |
|---|---|---|---|
| Text Embedding 3 Large | text-embedding-3-large | OpenAI | High-dimensional; strongest OpenAI . |
| Text Embedding 3 Small | text-embedding-3-small | OpenAI | Cost-optimized sibling; still current, chosen for cheap/high-volume retrieval. |
| Voyage 4 Large | voyage-4-large | Voyage | Highest-quality Voyage retrieval embedding. |
| Voyage 4 | voyage-4 | Voyage | Balanced Voyage embedding. |
| Voyage 4 Lite | voyage-4-lite | Voyage | Lower cost / latency. |
| Voyage 4 Nano | voyage-4-nano | Voyage | Smallest / cheapest Voyage embedding. |
| Voyage Context 4 | voyage-context-4 | Voyage | Contextualized-chunk embedding. |
| Voyage Multimodal 3.5 | voyage-multimodal-3-5 | Voyage | Current Voyage multimodal model - text, image, and video embeddings in one space. |
| Cohere Embed v4 | embed-v4 | Cohere | Multimodal (text + image); Matryoshka dimensions (256-1536). Pair with rerank-3.5 or rerank-4.0. |
| Gemini Embedding 2 | gemini-embedding-2 | Google's current embedding model. |
Anti-pattern example IDs
Some docs deliberately show outdated model IDs as examples of what not to ship. These are exempt from the model-ID check because their staleness is the point of the example. This section is their allowlist - keep it in sync with the anti-pattern docs.
| ID | Where | Why exempt |
|---|---|---|
claude-3-sonnet | context-files/anti-patterns.mdx | Deliberately-stale model pin in the "auto-generated boilerplate nobody curates" anti-pattern. |