Unknown model gpt-5.6-luna

Potential duplicates detected. Please review them and close your issue if it is a duplicate.

• #37536

Powered by Codex Action

Seeing this on my end too when trying to spawn Luna subagents from Sol as the main thread on the macOS desktop Codex app. The way codex described it to me, for additional context:

Evidence:

Official Codex documentation lists Luna for the desktop app, CLI, IDE extension, and Codex cloud. Codex models
The official subagent guide explicitly recommends gpt-5.6-luna for fast, narrow, high-volume subagents. It also includes custom-agent examples that use Luna. Codex subagents
The configuration reference supports agents.default_subagent_model. Configuration reference
Your local configuration has no subagent-model restriction.
Your local model cache includes gpt-5.6-luna.
A direct local test with codex exec -m gpt-5.6-luna succeeded.

A subagent test in this task failed with:

Unknown model `gpt-5.6-luna` for spawn_agent.
Available models: gpt-5.6-sol, gpt-5.6-terra

I can reproduce this in Codex Desktop.

The main task can be switched to gpt-5.6-luna, but the multi-agent spawn_agent model allowlist only exposes:

• gpt-5.6-sol
• gpt-5.6-terra

An explicit spawn attempt using:

{
  "model": "gpt-5.6-luna",
  "reasoning_effort": "low",
  "fork_turns": "none"
}

fails before creating an agent with:

Unknown model `gpt-5.6-luna` for spawn_agent.
Available models: gpt-5.6-sol, gpt-5.6-terra

Switching the parent/main task itself to Luna does not update the subagent allowlist, and an inherited/default spawn does not expose enough model metadata to verify that it remained on Luna.

This appears to be an incomplete multi-agent model registry or rollout rather than a Luna capability limitation. The official model page lists Luna as supporting tools and reasoning through max: https://developers.openai.com/api/docs/models/gpt-5.6-luna

Temporary workaround

Codex Desktop's separate task/thread creation API does accept gpt-5.6-luna with max reasoning. A working approach is:

1. Create a new local/sidebar task explicitly configured as gpt-5.6-luna with max reasoning.
2. Use that Luna task as the coordinator.
3. Have the Luna coordinator call spawn_agent without a model override, allowing its hidden subagents to inherit the coordinator's default model.
4. Message and monitor the Luna coordinator from the original task.

This is less convenient than selecting Luna directly in spawn_agent, and inherited subagents still do not expose an exact model identifier, but it currently provides a practical path for Luna-based multi-agent work.

My workaround so far has been to ask the orchestrator model to spawn a codex exec process using luna, instead of a sub-agent. That seems to work fine, but you lose the UX benefit and probably a few others in the process.

I reproduced the same issue on Windows with VS Code extension 26.803.61601 (bundled Codex 0.147.0-alpha.6.5).

I investigated it further in #38107 and found:

• Standalone Codex CLI 0.147.0: gpt-5.6-luna subagent spawn succeeds
• VS Code bundled 0.147.0-alpha.6.5: Luna spawn fails with the same Unknown model error
• The bundled binary also fails when launched directly outside VS Code
• After updating to VS Code extension 26.810.41047 (bundled Codex 0.148.0-alpha.9), Luna subagent spawn succeeds

The newer bundled runtime includes the leaf-worker support from #36892, which appears consistent with the fix.

So at least for the Windows VS Code extension, this appears resolved in 26.810.41047 / 0.148.0-alpha.9.

Full A/B investigation: #38107

Product feedback from an affected user: this bug has a direct cost impact.

Luna is explicitly positioned as the faster/lower-cost model for narrow, high-volume subagent work, yet affected Codex builds expose Luna as an available model while rejecting it in spawn_agent from a Sol parent. That forces users to run Sol/Terra workers or use awkward codex exec workarounds, which defeats the main cost-control reason to use Luna in a multi-agent workflow.

Please treat subagent model availability as an end-to-end compatibility contract