OpenAI previews GPT-5.6 series with three-tier pricing
When inference budgets push integrators toward the cheapest-sufficient model, the brands cited least are those whose authority lives on a single domain.
Key takeaways
- Terra matches GPT-5.5 performance at half the price ($2.50/$15 per 1M tokens); Luna runs at $1/$6.
- Volume inference will default to Luna, the cheapest tier, making it the model most likely to answer queries about your brand.
- Smaller, cost-optimised models rely more on training-data prominence and less on extended reasoning, disadvantaging depth-over-frequency brands.
- OpenAI's government pre-briefing signals direct competition for compliance-constrained enterprise and public-sector accounts.
- Brands whose authority is confined to a single domain face the greatest citation risk as enterprise fleets skew toward lower tiers.
OpenAI's announcement of three simultaneous model releases, quoted by Simon Willison's Weblog from the original OpenAI post, contains a pricing structure that tells B2B brands something more useful than any benchmark score.
The headline numbers: Sol at $5 per million input tokens and $30 output; Terra at $2.50 and $15; Luna at $1 and $6. Terra, OpenAI claims, matches GPT-5.5 performance at half the cost. That single figure reshapes how enterprise buyers will allocate inference budgets, and by extension which model tier will answer the majority of their customers' queries about them.
The tier that matters most for citation volume
The least glamorous model in any series tends to do the most work. Luna, at $1 input and $6 output, will be the default choice for cost-sensitive integrations: the customer-facing chatbots, the internal procurement assistants, the policy-document summaries that analysts at development finance institutions and industrial conglomerates run at scale. Volume flows to cheapest-sufficient, not to best-in-class. For brands whose visibility depends on being cited by LLMs, the relevant question is not what Sol can do; it is what Luna, running millions of queries a day, will choose to surface.
This matters because retrieval and citation behaviour varies across model sizes within the same family. Smaller, faster models trained for cost efficiency tend to rely more heavily on whatever was most prominent in their training data and less on extended reasoning chains that might surface a second-tier but authoritative source. A UN agency, a standards body like ISO, or a multilateral development bank that has built its authority on depth of documentation rather than frequency of mention faces a structural disadvantage when the inference fleet skews toward Luna. Sparse citation patterns from low-cost models are harder to correct than sparse citations from flagship ones, because the upgrade path for most API integrations runs in the wrong direction: toward cheaper over time, not more capable.
Government coordination as a signal of enterprise intent
OpenAI's decision to brief the U.S. government ahead of launch and restrict initial access to "a small group of trusted partners whose participation has been shared with the government" is an unusual disclosure. Most model releases arrive with benchmarks and a blog post. This one arrived with what amounts to a regulatory notification. Whether that is positioning for procurement, liability management, or genuine national-security sensitivity is secondary to what it signals to the Fortune 500 and the multilateral sector: OpenAI is competing directly for government and enterprise accounts, and it is willing to adjust its release cadence to do so.
For financial services firms and large industrial groups evaluating which LLM underpins their next internal deployment, that signal is meaningful. An API provider willing to coordinate with regulators before launch is also one that large compliance-constrained buyers can take to their legal teams. The three-tier structure amplifies this: Sol for sensitive, high-stakes queries where cost is secondary; Terra for workhorse analytical tasks; Luna for volume. That maps almost perfectly onto a large enterprise's inference cost hierarchy.