Superpower Daily
Cheshire Academy Uses Traffic-Light Labels to Set AI Boundaries...OpenAI Pauses Frontier Training After Sandbox Breach, Warns of...Cloudflare Open-Sources Cloudflare OS, Putting AI-Built Work Ap...EverestLabs Puts AI Agents Near Recycling Controls, but Keeps O...Lexis+ with Protégé Adds Agents, Context and Office Files to Le...Geneva Rally Presses Hochul on 20-MW New-Project Permit Pause a...Cheshire Academy Uses Traffic-Light Labels to Set AI Boundaries...OpenAI Pauses Frontier Training After Sandbox Breach, Warns of...Cloudflare Open-Sources Cloudflare OS, Putting AI-Built Work Ap...EverestLabs Puts AI Agents Near Recycling Controls, but Keeps O...Lexis+ with Protégé Adds Agents, Context and Office Files to Le...Geneva Rally Presses Hochul on 20-MW New-Project Permit Pause a...
Modelspublished

OpenAI Lets Codex Route Smaller Tasks From Sol to Lower-Cost Luna Workers

The new routing lets developers set model choice and reasoning effort by subtask, but the savings depend on work that can be handed off with complete context.

By 2 min read
OpenAI Lets Codex Route Smaller Tasks From Sol to Lower-Cost Luna Workers
Sign in to save and react

Listen to this story

The audio brief

About 1:23
0:001:23
Read transcript
OpenAI is letting Codex route smaller coding tasks from GPT-5.6 Sol to GPT-5.6 Luna, creating a cheaper second tier inside one multi-agent workflow. Sol remains the planner: it breaks down the job, gives Luna a bounded assignment, and assembles the returned work. Luna is the execution layer, designed for contained tasks, and it cannot contact other agents or spawn more workers. The important detail is that this is opt-in. After updating the app, developers must explicitly tell Codex to use Luna. Otherwise, Sol continues creating workers with the parent model, the same reasoning setting, and the existing forked context. That default is described as the stronger-performing mode in evaluations, but it is also slower and uses more tokens. The handoff is where the engineering challenge sits. A Luna worker may not have the parent conversation, so its opening instruction needs to include everything required to finish the assignment. OpenAI recommends using “fork turns: none” when the parent context is unnecessary, and suggests keeping the workflow to roughly six to eight subagents for reliability, although that is not a hard limit. Developers can now choose the model and reasoning effort task by task, rather than for an entire session. The question is whether teams can isolate routine work cleanly enough to capture Luna’s savings without removing context the task actually needs.

Story brief

3 key points

OpenAI’s Codex Multi Agents v2 now supports task-level routing from GPT-5.6 Sol to GPT-5.6 Luna, giving developers a cheaper execution option without changing the default behavior. Sol still plans, instructs, and assembles results, while Luna handles self-contained assignments and cannot spawn or contact other agents. The feature is opt-in, limited to supported Codex models, and requires complete handoff prompts...

  1. 01

    Luna routing is available after an app update, but developers must explicitly prompt Codex to use it.

  2. 02

    By default, Sol continues spawning workers with the parent model, reasoning setting, and forked context.

  3. 03

    Recommended context isolation uses `fork_turns: none` plus a complete opening instruction for each Luna worker.

OpenAI has added cross-model delegation to Codex Multi Agents v2, allowing GPT-5.6 Sol to direct a job while assigning bounded tasks to GPT-5.6 Luna. It creates a model hierarchy within a coding workflow: Sol handles decomposition and result assembly, while Luna takes contained work as the faster, lower-cost worker. The routing is limited to supported Codex models, rather than outside providers.

The design is narrower than a free-form agent swarm. Luna workers cannot message other agents or spawn more workers. Sol retains the orchestration tools, leaving the parent responsible for breaking down the job, supplying instructions, and collecting results.

That puts Luna in a role suited to clearly scoped execution, while Sol holds the wider plan. The handoff must stand on its own: a Luna worker without the parent conversation needs an initial instruction containing everything required to finish its assignment.

An opt-in route, not the new default

The feature becomes available after an app update, but developers must prompt Codex to use a different model. Without that instruction, it continues spawning workers with the parent model, reasoning setting, and context-forking behavior. A Sol session therefore still produces Sol workers by default.

The default is presented as the higher-performing mode in evaluations, but it runs more slowly and uses more tokens. Luna is the GPT-5.6 family’s most cost-efficient model and is positioned as its fastest, lowest-cost option. Cross-model routing offers a second mode built around specialization and explicit budgets.

Context is now a design constraint

For workers that do not need the parent conversation, the recommended setup is fork_turns: none and a complete opening prompt. Guidance also advises against running more than six to eight subagents because of reliability concerns. Those are operating recommendations, not enforced limits.

A missing worker tier becomes usable

The change resolves a prior routing gap: Multi Agents v2 had accepted Sol and Terra subagents while rejecting Luna as an unknown model. Earlier behavior also exposed a mismatch in which the runtime could accept model and reasoning-effort fields that were absent from the spawn-tool definition shown to the parent.

Developers can now select a model and reasoning effort task by task instead of making that choice for an entire session. The practical test is whether routine work can be isolated without stripping away the context needed to complete it correctly.

Related coverage

Continue reading