Model Resolution

xcaffold supports a model field on agents that accepts either a tier alias (balanced, flagship, fast) or a literal model ID (e.g., gpt-5.5, claude-sonnet-4-6). Each provider's model resolver translates these into the correct output value.

Tier Aliases

Tier aliases let you write provider-agnostic .xcaf files. Each provider maps the alias to its recommended model for that tier.

Claude Code uses bare aliases (sonnet, opus, haiku) that resolve at runtime to the latest version. This means the compiled output always targets the current model without needing resolver updates.

Antigravity 2.0 uses its own short aliases (flash, pro, pro-low, sonnet-thinking, opus-thinking, gpt-oss) rather than the generic tiers — see Antigravity 2.0 Models below.

Last verified: 2026-06-09 against official provider documentation. Tier mappings are defaults. Override them per-provider using model-tier-<alias> entries in project.<provider>.vars, or use agent.<provider>.xcaf overrides for individual agents.

Literal Model IDs (Pass-Through)

When the model field contains a literal model ID instead of a tier alias, each provider checks whether the ID matches a known prefix family and passes it through unchanged.

Pass-through is case-insensitive. Claude-Sonnet-4-6 resolves to claude-sonnet-4-6.

Unrecognized model IDs (no matching prefix) produce an AGENT_MODEL_UNMAPPED warning and are omitted from output.

Antigravity 2.0 Models

Antigravity 2.0 supports multi-vendor model selection — Gemini, Claude, and GPT-OSS models are all valid targets.

Full model IDs accepted unchanged: gemini-3.5-flash, gemini-3.1-pro-high, gemini-3.1-pro-low, gemini-3-flash, claude-sonnet-4-6-thinking, claude-opus-4-6-thinking, gpt-oss-120b, nano-banana-2 (image generation, UI-only).

Default model: gemini-3.5-flash.

Deprecated: the antigravity (v1) target does not compile the model field at all. New projects should target antigravity2. See Supported Providers.

Fidelity Notes

Per-Provider Override

To set a different model for a specific provider without changing the base .xcaf:

model: gpt-5.5

Or use variables:

model: ${var.model-deep}

model-deep = gpt-5.5

model-deep = claude-opus-4-7

User-Configurable Tier Mapping

To override what a tier alias resolves to for a specific provider, add model-tier-<alias> entries to the target's variable file:

xcaf/project.cursor.vars:

model-tier-balanced = composer-2.5-custom
model-tier-flagship = gpt-5.5
model-tier-fast = composer-2.5-turbo

When present, these override the built-in alias map for the compilation target. Any agent using model: balanced compiles to composer-2.5-custom on Cursor instead of the default. When absent, the built-in defaults apply.

This is useful when a team wants to standardize model selection across all agents without per-agent overrides or variable references.

Tier overrides apply only to the three canonical aliases. Literal model IDs and provider-native aliases are unaffected.

Cursor Model Catalog

Featured models available in Cursor as of 2026-06-09:

Source: Cursor Docs — Models (mined 2026-06-09). Model availability and slugs may change. Use literal pass-through for models not listed here.