GPT-5.6 Sol, Terra and Luna: What OpenAI's Tiers Mean for Enterprise AI

OpenAI made GPT-5.6 generally available on Thursday, after a delay requested by the US government, and split it into three tiers with very different prices: Sol, Terra and Luna. The same day it launched ChatGPT Work, an office agent built on the new models, and Microsoft made GPT-5.6 the preferred model in Microsoft 365 Copilot. For enterprise leaders, the release matters less for any single benchmark than for what it signals: model choice is becoming a cost-management decision made workload by workload.

What did OpenAI release with GPT-5.6?

GPT-5.6 rolled out across ChatGPT, Codex and the API on July 9, with full availability expected within a day, Nerova reported. The three tiers and their API prices per million tokens are:

• Sol, the flagship: $5 input, $30 output.
• Terra, the balanced option: $2.50 input, $15 output.
• Luna, the fastest and cheapest: $1 input, $6 output.

Luna's list price is one-fifth of Sol's. When OpenAI previewed the family in late June, it said Terra matches GPT-5.5-level performance at about half the cost, The National CIO Review reported. A new "ultra" setting coordinates several agents in parallel on the hardest tasks, spending more tokens to finish faster, according to Resultsense.

Inside ChatGPT, access depends on plan. Plus, Business and Enterprise users can use Sol at medium effort and above, while Pro and Enterprise users also get Sol Pro for long, complex work, according to The D(AI)ly Brief.

Why did the US government delay GPT-5.6?

The general release followed a pause. OpenAI held back the model in June at the US government's request over concerns that it could speed up sophisticated cyberattacks, and limited it to vetted partners, Resultsense reported, citing Reuters and Axios. Axios reported that the administration approved a broad rollout after further testing and meetings with officials.

OpenAI says the model stays below the "Cyber Critical" threshold in its Preparedness Framework and was red-teamed with more than 700,000 A100-equivalent GPU hours before release, according to The National CIO Review. Some requests touching cybersecurity or biological research may get extra checks or slower handling, Nerova noted.

This is a new kind of release risk for enterprise roadmaps. Teams that planned around a frontier model's launch date now have to consider that government review can move it, and that some use cases, such as security tooling, may face extra friction even after launch.

How does Microsoft Copilot change the picture?

Distribution may matter more than benchmarks. GPT-5.6 became the preferred model in Microsoft 365 Copilot on release day, inside Word, Excel, PowerPoint, Copilot Chat and Microsoft's Cowork agent, Resultsense reported. Many enterprises will meet the model through software they already license, not through a deliberate procurement decision.

ChatGPT Work, launched alongside, combines ChatGPT with the Codex coding tool so non-coders can build documents, presentations and websites from plain instructions, Resultsense reported, citing Reuters. It rolls out first to Pro, Enterprise and Edu users and competes with Anthropic's Claude Cowork and Microsoft's Copilot Cowork. That puts OpenAI in the odd position of selling a product that competes with an agent built by its biggest distribution partner.

Is a cheaper tier good enough for enterprise work?

This is where the evidence is thinnest. OpenAI's pitch centers on performance per dollar, and it claims Terra and Luna match or beat rivals at a fraction of the cost. An analyst quoted by Reuters found the smallest model could complete a task about as well as the largest at roughly one-fifth the price, according to Resultsense. One test is not a procurement standard, and vendor benchmarks rarely reflect a company's own documents, data quality and edge cases.

The skeptical view is that tiering shifts work onto buyers. Someone now has to decide which tier each workflow gets, monitor whether the cheaper tier is failing quietly, and revisit those decisions when prices or models change. The optimistic view is that tiering finally lets enterprises match cost to value, using the flagship only where it earns its price.

Both views point to the same requirement: enterprises need their own evaluation sets and cost tracking. Amey Desai, head of AI at Nexla, closed a AI Connect panel on scalable RAG systems by urging teams to "calculate the cost per question." That metric becomes more useful when the same question can be routed to three prices. Our coverage of building custom AI benchmarks explains how teams test models against their own work.

There is also a counterweight to deeper dependence on one provider. Speaking at a AI Connect webinar on Wednesday, Mistral AI's Aaron Delp made the case for enterprises controlling their models and deployment: "This isn't renting AI. This is owning your AI," as covered in our report on Mistral AI Studio.

What should enterprise leaders do now?

• Map workloads to tiers. Start high-volume, well-defined tasks on Luna or Terra and reserve Sol for work where errors are costly.
• Build a routing and evaluation layer. Keep prompts and evaluation sets portable so a tier or provider can be swapped without rewriting applications.
• Check what arrived through Copilot. Confirm which model your Microsoft 365 tenants now use and update data and usage policies.
• Plan for release risk. Treat frontier model availability as uncertain, especially for security-related use cases.

The bigger shift is that price lists are now a strategic input. Enterprises that can measure quality and cost per task will benefit from three tiers. Those that cannot will likely default to the most expensive one.

Frequently asked questions

What are the GPT-5.6 tiers and prices?

Sol costs $5 input and $30 output per million tokens, Terra $2.50 and $15, and Luna $1 and $6.

Why was GPT-5.6 delayed?

OpenAI held it back in June at the US government's request over concerns it could accelerate cyberattacks; Axios reported the administration later approved a broad rollout.

Which GPT-5.6 tier should enterprises use?

Start high-volume, well-defined tasks on Luna or Terra, reserve Sol for costly-error work, and validate with your own evaluation sets and cost-per-task tracking.

Filed under: AI Strategy & ROI, Agentic AI, AI Governance & Trust