OpenAI lets GPT-5.6 Sol delegate grunt work to cheaper Luna agents

Why it matters

Cross-model delegation gives Codex a direct cost and latency control: expensive reasoning can stay with the orchestrator while faster models handle clearly bounded work.

OpenAI has shipped cross-model delegation for Codex's Multi Agents v2 system, allowing a model to assign work to any supported model, including GPT-5.6 Luna, Eric Provencher (@pvncher) said in a thread on X on August 15th.

The release gives Codex a practical model hierarchy: a capable model such as GPT-5.6 Sol can remain the orchestrator while sending narrowly defined tasks to a faster worker. OpenAI describes Luna as the fastest and lowest-cost model in the GPT-5.6 family, making the routing useful for controlling latency and consumption during agent-heavy coding jobs.

Provencher joined OpenAI in developer relations after building Repo Prompt, a native macOS tool for curating codebase context and coordinating coding agents. Before working on Repo Prompt full time, he spent five years at Unity. Repo Prompt became open source on June 13th, with its Community Edition continuing as an agent-orchestration project.

That background maps directly onto the feature Provencher is shipping at OpenAI. Repo Prompt was built around separating planning, context gathering and implementation across specialized agents. Multi Agents v2 now gives Codex a similar division of labor within OpenAI's model lineup.

Sol directs, Luna executes

Provencher described Luna workers as "pure sub agents." They do not receive the same cross-agent coordination capabilities as peer agents and cannot message or spawn additional agents. Sol's orchestration tools remain available, leaving the parent model responsible for decomposing the job, issuing instructions and collecting results.

The distinction matters for reliability. A Luna worker is best suited to straightforward, bounded work with an explicit starting prompt, Provencher said. He recommended using fork_turns: none when the worker does not need the parent conversation and ensuring that its initial instructions contain everything required to finish the task.

That setup lets a developer use Sol for decisions that depend on broad context while reserving Luna for mechanical implementation, searches or other isolated work. The parent remains accountable for assembling the result rather than allowing an expanding tree of workers to coordinate among themselves.

Provencher said the feature works without additional configuration after users update the app, although the Codex app can receive functionality before the corresponding version bump. Users must currently prompt Codex to route work this way. The default behavior still spawns copies using the parent's model, reasoning effort and context-forking settings.

That default limits accidental changes in behavior, but it also means the cost benefit is not automatic. A Sol session will continue creating Sol workers unless the prompt tells the orchestrator to choose a different model and reasoning level.

The release closes a visible routing gap

Cross-model delegation addresses a limitation Codex users had documented throughout July. In one GitHub issue filed on July 22nd, a user running Codex CLI 0.145.0 reported that Multi Agents v2 accepted Sol and Terra subagents while rejecting Luna as an unknown model. A separate issue about the spawn schema found that the runtime could accept model and reasoning-effort fields even when those controls were absent from the tool definition shown to the parent agent.

Those failures undercut the main economic argument for multi-model orchestration. If every worker inherits the most capable parent model, simple tasks consume the same model tier as planning and debugging. Exposing model and effort selection lets the orchestrator allocate capability task by task.

Provencher said OpenAI took additional time to make the routing reliable. He also advised against running more than six to eight subagents, setting a practical ceiling below the large swarms often used in agent demonstrations.

Codex's existing default remains the higher-performing option in OpenAI's evaluations, according to Provencher: the parent creates workers with the same model, reasoning effort and forked context. He said that configuration runs more slowly and consumes more tokens. The new routing gives developers a second operating mode built around specialization and explicit budgets.

The scope remains OpenAI's supported Codex models. Provencher's announcement does not establish general routing across outside model providers. Within the GPT-5.6 family, however, Codex can now separate the model making the plan from the model carrying it out, turning model choice into an orchestration decision rather than a session-wide setting.