Models

Guard provides the Claude models listed below. Choose the exact model and a compatible API; see billing for rates and charging rules.

Free-form spellings are mapped to canonical ids. Claude Code suffixes like claude-opus-4-8[1m] are stripped before the upstream.

For Kimi, the display name and API ID are separate. Copy the API ID into the model field without changes.

The context window covers the whole request together with the history and the model's answer. For k3 and kimi-for-coding it is 1,048,576 tokens; for k3-256k and kimi-for-coding-highspeed it is 262,144. What to take:

Kimi has no free-form spellings; the id is matched exactly. K3, kimi-k3 or Kimi K3 return 404 model_not_found before any charge, so copy the id from the table as it is.

Two families, 18 exact IDs in total: two Composer, eight Grok 4.7, and eight Grok 4.6. Grok effort and the Fast variant are part of the ID; a separate field does not switch them.

For Grok 4.7, the effort level is the -low, -medium, -high, or -xhigh ending, and a final -fast chooses the faster tariff. All eight ids use a 256K window. Grok 4.6 keeps its previous eight ids. Effort itself does not change the rate.

Grok 4.7 has eight ids: four effort levels, each in standard and Fast variants. Grok 4.6 keeps four standard and four Fast ids.

Composer has two ids: composer-2.5 and composer-2.5-fast. There are no effort levels, images are not accepted, and the context window is 200,000 tokens. The -fast variant is a separate tariff: the same answer arrives faster and costs more.

All 26 Composer/Grok ids are returned by the shared GET /v1/models. Copy the id into the model field without changes.

The seven Codex models are available through both namespaces, but their catalog shapes differ. Without a query, GET /codex/v1/models returns the OpenAI exact7 list, while shared GET /v1/models returns the common OpenAI catalog including exact7. client_version is asymmetric: only /codex/v1/models?client_version=… returns raw native exact7; /v1/models?client_version=… keeps the legacy Claude client catalog. A broad gpt-* pattern does not make an unknown model available.