Codex Multi-Agents v2: Sol Delegates Grunt Work to Cheaper Luna

By Deepak Bagada, CEO at SaaSNext & Principal AI Architect.

Multi-agent coding has an efficiency problem that every team using it eventually hits: the orchestrator and its workers all run on the same model, so the tokens spent on repetitive grunt work cost exactly as much as the tokens spent on deep reasoning. OpenAI's Codex Multi-Agents v2, shipped on August 15, 2026 and announced by Eric Provencher on X, attacks that problem directly with cross-model delegation. An orchestrator agent named Sol can now hand narrowly-defined tasks to Luna, the fastest and lowest-cost model in the GPT-5.6 family. The result is a simple but profound idea: model choice is now an orchestration decision.

Cross-model delegation: the announcement

The August 15, 2026 release makes delegation an explicit, first-class mechanism. Instead of every agent in a Codex session being a clone of the same model, the orchestrator can dispatch a task to a cheaper, faster model that is purpose-built for it. Luna is described as a pure sub agent: it cannot message other agents and cannot spawn new ones. That constraint is a feature. It keeps the delegation graph flat — Sol is the only coordinator, and every Luna worker is confined to the task it was given. No runaway chains of agents spawning agents, no unbounded tree of nested reasoning.

The design reflects a lesson the industry has been learning the hard way: autonomous agent graphs become unmanageable the moment agents can freely create other agents. Luna's crippling is deliberate containment.

Luna: the fastest, cheapest, and deliberately crippled

Luna is not a smaller, dumber Sol. It is a member of the GPT-5.6 family tuned for speed and cost, and its limits are part of its value proposition:

• Fast and cheap per token — the right economics for high-volume, low-complexity work.
• No messaging — it cannot talk to sibling agents, only answer Sol.
• No spawning — it cannot create new agents, so delegation stays bounded.

If Sol is the senior engineer who holds the plan and the review bar, Luna is the disciplined junior who executes one well-specified task and hands the result back. That division of labor is exactly what makes the cost model interesting.

Routing still needs prompting

The headline caveat is that routing is not automatic. Cross-model delegation requires prompting — you must explicitly ask Sol to hand a task to Luna — and the default remains same-model delegation. Provencher's own guidance reinforces the practical ceiling: keep sub-agent counts around 6-8. Past that point, coordination overhead and context churn grow faster than parallelism gains, and the orchestrator spends more tokens managing workers than workers spend doing work.

The July 22, 2026 GitHub issues are a useful reminder that this is a real migration, not a flip of a switch. Teams reported Luna being rejected as an unknown model when a workflow referenced it, and a spawn-schema change breaking existing configurations. The lesson: if you depend on multi-agent workflows, treat v2 as a platform upgrade with a changelog, not an invisible backend change.

Sol versus Luna: where the tokens go

Read the table as an allocation problem. If Sol plans and reviews but never touches the repetitive token-heavy work, and Luna absorbs the grunt loops, the blended cost per completed task drops while the quality ceiling is preserved. The economics work because the two models price their tokens differently for the same family.

The token-cost math

A worked example makes the leverage concrete. Imagine a session where 70% of tokens go to narrow sub-tasks — reading files, searching symbol usage, formatting output, retrying failed calls — and 30% goes to Sol's planning and review. If Luna's per-token price is a fraction of Sol's, the bill for that 70% shrinks proportionally. Even with a 4x price gap between tiers, routing the grunt work to Luna could cut overall token spend by roughly half, because the expensive tokens are now concentrated where the hard reasoning actually happens.

The counter-case matters too. If your workload is mostly deep reasoning with very little delegatable grunt work, the savings shrink because Sol dominates the bill regardless of what Luna costs. The takeaway is managerial: before enabling cross-model delegation, measure your token mix. Model choice in a multi-agent system is a budget-allocation decision, and the allocation should follow the work.

Putting delegation to work

The practical playbook for cross-model delegation is short, and it starts before you write a prompt. First, instrument the session: log tokens per agent and per task class so you know what is actually consuming your budget, not what you assume is. Second, find the narrow, repeatable tasks — file search, symbol resolution, output formatting, error-loop retries — and extract them into their own explicit prompts that a worker can execute without needing Sol's context. Third, keep Sol's context lean. The orchestrator's reasoning budget is the expensive part, so feed it plans and summaries rather than raw dumps; the more a session reads like a manager and writes like a delegation, the better the cost profile. Fourth, enforce the fan-out ceiling in your own tooling rather than relying on discipline. A hard cap of six to eight active workers per session is a config value, not a vibe, and it protects you from the coordination tax that silently eats the parallelism gains. Finally, treat the July 22 breakage as a warning to pin versions: cross-model delegation changed the spawn schema and the model registry, so pin your agent runtime and test delegation paths in CI before you roll them out to the whole team. None of this is exotic — it is ordinary cost and change management, applied to a system that finally lets you pay for capability instead of token volume.

Frequently Asked Questions

Can Luna spawn or message other agents?

No. Luna is deliberately a pure sub agent with no messaging and no spawning. That constraint keeps the delegation graph flat, prevents runaway agent chains, and limits the blast radius of any single worker to the task Sol assigned.

Is cross-model routing automatic?

No. Routing Luna tasks requires prompting; the default remains same-model delegation. Teams that want the cost benefit must build the delegation into their task definitions and prompts.

Why does Provencher recommend 6-8 sub-agents?

Coordination overhead and context churn grow quickly past that point. Each worker adds message traffic and state Sol must track, so beyond roughly 6-8 workers the orchestration cost can outweigh the parallelism benefit.

What were the July 22, 2026 GitHub issues about?

Teams hit two transition pains: Luna rejected as an unknown model when referenced by workflows, and a spawn-schema change breaking existing multi-agent configurations. Both are migration costs of the v2 platform change.

Does cross-model delegation always save money?

Only when a meaningful share of tokens is spent on narrow, repetitive work that can be routed to Luna. If the workload is mostly deep reasoning, Sol dominates the bill and the savings shrink. Measure your token mix before you delegate.

Closing thoughts

Codex Multi-Agents v2 makes a point that generalizes beyond OpenAI: in a multi-agent system, the model is not a fixed substrate, it is an allocation variable. Sol and Luna are a rehearsal of a future where every agent fleet is a portfolio — frontier reasoning where it matters, cheap high-throughput tokens everywhere else. The constraints that make it work — pure sub agents, prompting-based routing, hard caps on fan-out — are the same constraints that make the system auditable and affordable. Teams that treat delegation as a cost-allocation decision will get the leverage; teams that leave it on default will pay Sol prices for Luna work. For more on building and costing agent fleets, see the workflows library and track agent-model economics in the latest AI news.