On 22 September 2026 Naomi Moneypenny, a Partner Group Product Manager at Microsoft, announced general availability of GPT-6 Sol and GPT-6 Luna in Microsoft Foundry. The post does not reopen the GPT-6 family’s debut—already covered when OpenAI launched Sol and Luna—but the Azure layer: regions, data zones, provisioned throughput and priority processing for production agents.
Foundry presents Astra, Sol and Luna as a fleet with roles. Astra for demanding work; Sol and Luna to scale high-volume tasks with less “noise” and more ability to complete full tasks with agents, in Microsoft’s language. The commercial promise is to move from experimentation to production without changing control plane.
On deployment, Standard is available for Astra, Sol and Luna across 28 global regions and U.S. and EU Data Zones. Provisioned Throughput covers Astra and Sol in global regions and those Data Zones. Priority Processing, aimed at latency, is offered for Sol in global regions and the U.S. Data Zone. The map matters to compliance officers as much as to SREs.
Microsoft notes that OpenAI still launches first on Azure and cites sophisticated customers such as Manus choosing Foundry. The competitive subtext is direct versus Bedrock—where Grok 4.7 arrived on 28 September—and versus Google Cloud: the war is no longer only which model, but which residency, logging and quota controls travel with it.
For agent architects, Sol/Luna in GA unlocks SLAs and capacity patterns previews do not offer. They can reserve throughput, pin data to the EU and combine Priority Processing on interactive paths with Standard on batch. That is unglamorous plumbing and very expensive to improvise on top of a consumer API.
The announcement rhymes with the 25 September Copilot Home/Code/Autopilot story—covered on the 29th—but operates on another layer: Foundry is the model plane for builders; Copilot is the employee surface. Firms can run Sol in internal Foundry agents while end users see Autopilot without them being the same product.
The adversarial-distillation news adds urgency. THE DECODER documented that reasoning protections reached Azure late after OpenAI’s own APIs were closed. Foundry GA does not absolve Microsoft of safeguard parity; the more production traffic Sol/Luna attracts, the more visible any hyperscaler security-patch lag becomes.
Buyers should demand in contracts: a maximum propagation time for lab security mitigations into each Data Zone; an inventory of whether encrypted reasoning and anti-distillation are active; and a right to suspend a model ID if a partner evaluation reproduces extraction. GA without those clauses is token availability, not trust.
On pricing and quotas, Provisioned Throughput changes the economics of always-on agents: it is no longer only pay-as-you-go but reserved capacity. CFO and platform engineering must model Priority versus Standard routes; the classic mistake is paying Priority for overnight jobs that tolerate Flex-like latency.
Foundry’s model ecosystem—OpenAI, xAI, Meta, Mistral, DeepSeek, Black Forest Labs, per Moneypenny’s own product remit—turns Azure into a multi-vendor mall. Sol/Luna GA is an OpenAI anchor inside that mall, not the catalogue’s close.
Risks include naming confusion between GPT-6 Sol “the model” and Sol “the Foundry SKU”; dependence on a single cloud for critical agents; and the European perception that an EU Data Zone is not sovereignty if the control plane still sits elsewhere. Public procurers will ask for diagrams, not slogans.
Neural Pulse’s summary: on 22 September Microsoft industrialised GPT-6 Sol and Luna for production in Foundry. In the week of 1 October, with Bedrock signing Grok and DeepMind opening Argon to defenders, Azure’s differentiator will be how fast it matches frontier safeguards in every region—not merely the region count on the blog.
