OpenAI's Multi-Agent v2 turns tiered-model cost arbitrage into a supported architecture

OpenAI has shipped Multi-Agent v2, which lets its GPT-5.6 Sol and Terra models delegate tasks directly to Luna, the speed-optimized variant that a version-classification mismatch had kept out of delegation ^[1][2]. The consequence is not a new capability so much as a supported one: splitting a workload across price tiers inside a single agent system stops being a workaround and becomes an architecture you can build on ^[4][5].

The mechanics are mundane, which is the point. Luna was classified as a Multi-Agent v1 model while Sol and Terra were v2, so a v2 primary agent could not hand it work ^[2]. Under certain conditions delegation simply failed, leaving developers to route everything through one tier or build awkward workarounds ^[4]. A Sol or Terra primary can now spin up Luna as a subagent, assign a task, and receive the output inside a single orchestrated workflow ^[5]. Note what that implies: the break was internal to one model generation, since all three variants are GPT-5.6 ^[13].

The tiering is the reason anyone cares. Sol takes heavy reasoning, Terra sits in the middle trading capability against cost, and Luna is built for high-volume work where latency and price matter more than raw capability ^[6]. Cryptobriefing's example is a customer service stack that keeps nuanced complaint resolution on Sol and sends FAQ responses to Luna at a fraction of the cost and latency, an arrangement the publication says was not reliably possible before the update ^[7][8].

Two cautions for anyone about to redraw a cost model on this. First, the report carries no per-token prices for any of the three tiers, so the arbitrage is directional rather than quantified; the saving depends entirely on what share of your calls are genuinely safe to push down a tier ^[14]. Second, the router now becomes part of your failure surface. A misclassified request sends work that needs multi-step logic to the model chosen for speed, and the visible symptom is a wrong answer delivered quickly, not an error.

The provenance around the fix is thinner than the fix itself. The report says the mismatch was actively disrupting project workflows for developers who had assumed full interoperability across the family ^[9], that community feedback from developer forums suggests the frustration was widespread enough to make this a priority ^[10], and that OpenAI reportedly conducted reliability testing before shipping ^[11]. Those three are hedged in the source, and the last is worth treating as unconfirmed until OpenAI says it in its own words.

What to watch is the tooling lag and the next release. Developers have flagged that Codex still needs additional configuration adjustments to use v2 multi-agent functionality fully, so the core delegation pipeline works while the surrounding ecosystem has not caught up ^[12]. The structural question is whether version classification is now a documented contract or an internal detail that can shift again. A cost structure that depends on which multi-agent version a vendor assigns to its cheapest model is a cost structure exposed to that vendor's release schedule, and the frustration this time ran for weeks before it was fixed ^[3].