GPT-5.6 launches: what Luna, Terra, and Sol mean for LLM visibility

Three model sizes, one knowledge cutoff date, and a pricing structure that reframes how B2B brands should think about LLM visibility across the entire OpenAI ecosystem.

Simon Willison's Weblog reports that OpenAI's GPT-5.6 family, launched to general availability on 9 July 2026, ships in three tiers named Luna, Terra, and Sol, priced at $1/$6, $2.50/$15, and $5/$30 per million input/output tokens respectively. All three share a February 16th 2026 knowledge cutoff, a one-million-token context window, and a 128,000-token maximum output. OpenAI's headline benchmark claim is long-running agentic performance: on Agents' Last Exam, all three models reportedly outperform Claude Fable 5.

The cutoff date is the detail that matters most for brand visibility, and it is the one most commentators will skip past.

What a shared cutoff date does to citation economics

A single February 2026 cutoff applied across three price points means the same knowledge base now serves the casual consumer using Luna and the enterprise deploying Sol for complex agentic workflows. For a multilateral institution, a policy think-tank, or an industrial group trying to ensure its research appears in LLM answers, that uniformity is both a constraint and an opportunity. Any content published, updated, or indexed after February 16th 2026 is absent from all three models until the next training cycle. There is no premium tier that saw more recent data; Sol costs five times Luna per token and knows nothing Sol does not.

This matters because the distribution of GPT-5.6 usage will not be uniform across tiers. Luna, at $1 per million input tokens, will almost certainly handle the high-volume, lower-complexity queries: customer service automations, quick-answer interfaces, internal knowledge assistants at scale. Sol will run the agentic, multi-step work where citation quality and source authority are load-bearing. A brand that appears in Sol's training data and benchmark evaluations gains visibility in the decisions that matter most; one that missed the cutoff loses ground precisely where stakes are highest.

The pricing comparison with Anthropic is instructive. Claude Fable 5 sits at $10/$50 per million tokens, double Sol's input price. Claude Opus runs $5/$25, matching Sol on inputs. If OpenAI's agentic benchmark claims hold under independent scrutiny, enterprises will face genuine substitution pressure toward GPT-5.6 Sol for long-horizon tasks. That shifts citation weight toward OpenAI's training corpus and away from Anthropic's. Brands whose authority is better represented in one vendor's data than another's will see their LLM visibility move accordingly, through no action of their own.

The tiered model problem for content strategy

The three-size architecture creates a targeting problem that did not exist when there was effectively one GPT-4 class model. A financial services firm or UN agency that invests in structured, citation-worthy content needs to ask which tier of GPT-5.6 its likely interlocutors are using. An analyst at a major bank using an enterprise Sol deployment asking about sovereign debt restructuring will receive a different quality of answer than a junior researcher using a Luna-powered assistant at a smaller institution. The underlying knowledge base is identical; the reasoning budget is not. Sol will synthesise longer source chains; Luna will compress. Content designed to survive compression, meaning short declarative claims with named figures and institutional provenance, performs better across all three tiers than content that relies on context to carry its argument.