GLM-52 897 —
GPT-56SC 873 —
CL-OP5X 865 -0.9%
GROK-46H 865 -0.9%
GEM-37FH 865 -0.9%
GPT-56T 861 —
GLM-5 856 —
MUSE-SPK 841 —
QWEN-38X 824 -2.3%
GPT-6A 820 —
KIMI-K3X 810 -1%
CL-FAB5H 787 -0.9%
CL-OP5H 764 -0.9%
CL-OP46H 742 -0.9%
CL-OP47H 733 -1.1%
GEM-38FH 676 -1%
CL-OP47 585 -0.7%
INKL 531 —
CL-OP46 496 -0.2%
CL-OP48 490 -0.2%
GLM-52 897 —
GPT-56SC 873 —
CL-OP5X 865 -0.9%
GROK-46H 865 -0.9%
GEM-37FH 865 -0.9%
GPT-56T 861 —
GLM-5 856 —
MUSE-SPK 841 —
QWEN-38X 824 -2.3%
GPT-6A 820 —
KIMI-K3X 810 -1%
CL-FAB5H 787 -0.9%
CL-OP5H 764 -0.9%
CL-OP46H 742 -0.9%
CL-OP47H 733 -1.1%
GEM-38FH 676 -1%
CL-OP47 585 -0.7%
INKL 531 —
CL-OP46 496 -0.2%
CL-OP48 490 -0.2%
← Back to feed

OpenAI Codex Adds Multi-Agent V2: Sol Can Now Delegate Tasks to Luna Subagents to Cut Credit Burn

OpenAI added Multi-Agent V2 support to Codex this week. The update lets GPT-5.6 Sol spawn Luna subagents for bounded subtasks rather than resolving everything at Sol’s token cost. Luna is the practical delegation target given its pricing tier, but the feature supports any model in the Multi-Agent V2 registry.

How Delegation Works

Sol acts as orchestrator. For well-defined subtasks — test generation, documentation passes, constrained code review — it can spawn a Luna agent under the Multi-Agent V2 harness. Community-documented configuration guidance:

  1. Prompt the orchestrator explicitly. Sol does not self-demote by default. The system prompt must specify that Sol should select subagent capability based on required effort. Without this, Sol handles everything.
  2. Fork conversation context before passing to Luna. Luna without a fork inherits Sol’s full reasoning chain, which inflates context size and can confuse scope boundaries. Fork turns: none is the recommended setting for Luna subagents on straightforward, well-defined tasks.
  3. Luna is a pure subagent. It cannot act as a peer orchestrator or initiate further delegation. All coordination flows upward to Sol.

Credit Economics

Luna sits below Sol on both capability and cost. Tasks that are “straightforward, well-defined, and bounded” — the documented Luna threshold — consume credits at Luna’s rate when delegated, not Sol’s. For workflows where 30–50% of tasks fall into this category, the effective credit burn decreases noticeably. High-volume boilerplate work (tests, docstrings, format passes) sees the largest gains.

The August Codex efficiency improvement (previously reported as +18% session duration) was a separate backend fix. Multi-Agent V2 delegation is an additive credit-reduction path on top of that.

Model Interop Beyond Luna

Sol can delegate to any model in the Multi-Agent V2 registry, not exclusively Luna. This enables domain-specialized models for specific subtasks while Sol manages the task graph. The architecture is extensible — OpenAI has not published the full registry, but GPT-5.6 Terra and supported third-party models are accessible where API access is available.

Current Codex Context

  • Codex reached 8M users as of July 2026; usage limits have reset five times since launch.
  • GPT-5.6 Sol: 96.2% SWE-bench Verified, Terminal-Bench 2.1 88.8%.
  • GPT-5.6 Luna: 93.0% SWE-bench Verified, Terminal-Bench 2.1 84.7%.

The capability gap between Sol and Luna at the benchmark level is ~3 points on SWE-bench Verified. For bounded, low-ambiguity tasks, that gap is negligible in practice. Multi-Agent V2 targets exactly those cases.