OpenAI tunes GPT-5.6 Sol and opens Luna to free users
Sol's accuracy fixes and Luna's free-tier launch mean B2B brands now face two distinct LLM citation environments inside ChatGPT.
Key takeaways
- GPT-5.6 Sol has received accuracy and consistency fixes, making its citations more selective and harder to earn passively.
- GPT-5.6 Luna is now free to all users, making it the model most people encounter, and the most important for brand citation volume.
- Sol and Luna should be treated as distinct citation environments until evidence shows they behave identically.
- Institutions in financial services, policy, and the UN system face citation instability risks if Sol's consistency gaps have affected their prior visibility.
- Free-tier expansion means LLM citation strategy is no longer a premium-user concern; it is now the default audience.
OpenAI shipped two changes on the same day, and the pairing is more revealing than either update on its own. Per the OpenAI blog, GPT-5.6 Sol has received accuracy and consistency improvements, while GPT-5.6 Luna, previously gated behind paid tiers, is now available to free users with unlimited everyday chat access.
The Sol fixes address what OpenAI describes as gaps in accuracy and consistency. That framing is notable because Sol sits at the sharper, more reasoning-intensive end of the 5.6 family. When a model used for complex queries produces inconsistent answers, the downstream effect is not just a worse user experience; it is citation instability. A brand or institution that earns a citation in one Sol response may not appear in the next one on the same prompt. For enterprises in financial services or the UN system, where a model's answer to a governance or risk query carries reputational weight, inconsistency in citations is a structural problem, not an edge case.
The free-tier expansion is the bigger commercial signal
Luna's migration to the free tier matters more for brand visibility than the Sol accuracy patch. Free users represent the largest and fastest-growing share of ChatGPT's user base. If Luna is the model that answers most of those queries, then the citation patterns it has learned, and the sources it trusts, will shape how hundreds of millions of people encounter authoritative information. A brand that has optimised its content for visibility in premium-tier model responses may find its citations thin in Luna's outputs if the two models have meaningfully different retrieval preferences or training emphases.
OpenAI has not published the architectural differences between Sol and Luna in granular detail. What is clear is that they are distinct enough to be deployed separately, with different use-case targeting. That distinction should prompt B2B content teams to test both, not assume that visibility in one transfers to the other.
The timing also reflects a competitive reality. Google's Gemini free tier has been expanding its capabilities, and Anthropic has made Claude accessible at lower price points. OpenAI's move to put Luna on the free tier is partly a retention play on user volume. For brands, the consequence is that the audience interacting with free-tier LLMs is no longer a secondary concern: it is the primary one.
Industrial groups, multilateral institutions, and policy bodies that have invested in technical publishing, white papers, and structured knowledge bases have a specific stake in how Luna handles authoritative sources. If Luna's training and retrieval behaviour favours the same high-trust publishers as Sol, that content investment carries over. If it does not, the brand-visibility gap between organisations with strong LLM citation profiles and those without will widen faster than most communications teams are currently modelling.
The practical implication: treat Sol and Luna as separate citation environments until there is evidence they behave identically. Run the same priority queries through both, record which sources each cites, and map where your organisation appears or is absent. The accuracy improvements to Sol may also mean that some sources previously cited as hedges against inconsistency will drop out as the model becomes more selective. Appearing in a more accurate Sol is worth more; appearing in an inaccurate one is worth less than it looked.