OpenAI launches GPT-6 in two tiers, Sol and Luna

OpenAI has never before shipped a flagship model with a sibling. GPT-6 arrives as two: Sol, the frontier model aimed at hard reasoning and enterprise workloads, and Luna, a cheaper, faster variant built for what OpenAI's blog calls "everyday work." The company frames this as a matter of cost and speed, letting developers choose the right tool for the job. That is true as far as it goes. It is also the least interesting part of the announcement.

The interesting part is that OpenAI has, for the first time, formalised a two-tier retrieval architecture at the model level. Sol and Luna will not necessarily draw on the same web indices, apply the same freshness weighting, or cite sources with the same rigour. OpenAI's own materials describe Luna as optimised for latency and cost, which in practice tends to mean lighter-weight retrieval and shorter context windows devoted to source material. Sol, doing the heavier reasoning, is more likely to be the model plugged into ChatGPT's enterprise and research-grade surfaces where citation quality actually gets scrutinised.

Two models, two audiences, two citation regimes

This matters because most brand-visibility work to date has assumed a single, unified "ChatGPT" to optimise for. That assumption was already shaky given routing between GPT-4o, o-series reasoning models, and mini variants. GPT-6's launch makes the fiction untenable. A financial-services firm whose content gets cited when a user hits Sol for a due-diligence question may simply not appear when a different user, or the same user on a cheaper plan, gets routed to Luna for a quick market summary. The visibility a brand has painstakingly built into one model's retrieval layer does not automatically transfer to the other.

OpenAI has not published the routing logic that decides who gets Sol and who gets Luna, and it is unlikely to. That opacity is the real story. Google at least talks publicly, if vaguely, about how AI Overviews selects sources. OpenAI's blog post is a capability pitch, not a technical disclosure, and it says nothing about differential indexing, source freshness, or whether Luna's answers will carry citations at all in the way Sol's do. For a UN agency or multilateral body whose comms teams have spent the past year getting their reports correctly attributed in ChatGPT answers, that is not a footnote. It is a warning that the citation behaviour they benchmarked against GPT-5 may quietly bifurcate under GPT-6, with no changelog explaining why traffic patterns shifted.

The tiering also has a pricing logic that maps uncomfortably well onto who gets seen. Enterprise API customers, the ones running high-value workloads such as regulatory analysis or investment research, are the natural users of Sol, the model more likely to be tuned for accuracy and sourcing over speed. Consumer-tier and cost-sensitive deployments gravitate to Luna. If citation fidelity tracks that same split, and there is no reason yet to assume otherwise, then the brands with the resources to be cited well in expensive, careful answers are not the same brands showing up in cheap, fast ones. That is a two-speed information economy, and it did not exist as an explicit product decision before this week.

None of this is unprecedented in the industry; Anthropic already runs Claude Opus, Sonnet, and Haiku as an explicit capability ladder, and Google has long split Gemini into Pro and Flash. What is new is that OpenAI, whose ChatGPT remains the reference point for most B2B brand-visibility audits, has now joined that pattern at its flagship. The single-model mental model that has underpinned a year of AI-search-optimisation advice from agencies, including some of the frameworks now standard in enterprise comms, is out of date as of this launch.

The practical consequence is that brand-visibility audits built around "does ChatGPT cite us" need a second axis: which ChatGPT. Industrial groups and financial institutions running visibility tracking should ask their vendors, or their own teams, whether Sol and Luna are being monitored separately, because a citation rate that looks stable in aggregate could be masking a Sol gain offset by a Luna collapse, or the reverse. Aggregate numbers on a bifurcated model are not a measurement. They are an average of two different products pretending to be one.