GPT-5.6 Family Launches: Luna, Terra, and Sol Hit General Availability

OpenAI's GPT-5.6 family reached general availability today with three named tiers: Luna (smallest), Terra (mid), and Sol (largest), each targeting a distinct cost-performance band. The launch also makes GPT-5.6 the default model inside Microsoft 365 Copilot, pushing it into Word, Excel, PowerPoint, and Copilot Chat for enterprise users immediately.

Why it matters

Three-tier families are now the industry norm, but the GPT-5.6 pricing lands competitively against the current Anthropic stack. Luna at $1/$6, Terra at $2.50/$15, and Sol at $5/$30 per million input/output tokens positions Sol below Claude Fable 5 ($10/$50) and roughly on par with Claude Opus on input, though output pricing differs. As Simon Willison notes, raw per-token price is increasingly misleading once reasoning token counts vary by task, so effective cost depends heavily on workload.

Price-per-million tokens doesn't tell us much now that the number of reasoning tokens can differ so much between models.

For prompt engineering teams, the practical question is not just sticker price but reasoning efficiency: a cheaper model that burns 10x the reasoning tokens on a complex task is not actually cheaper.

What changes in practice

• Luna becomes the obvious default for high-volume, low-complexity tasks: classification, routing, extraction, short-form generation.
• Terra fills the mid-tier gap for moderate reasoning tasks where GPT-4o-class capability was previously the ceiling.
• Sol competes directly at the frontier tier, now priced to make head-to-head benchmarking against Claude Opus and Gemini Ultra practical for budget-conscious teams.
• Microsoft 365 Copilot integrations now run on GPT-5.6 by default, meaning enterprise prompt templates tuned for older models may need behavioral re-validation.
• Reasoning token variance means cost estimation requires profiling, not just multiplying token counts by rate card.

How to use it

1. Benchmark your actual workload first. Run your existing eval suite against Luna, Terra, and Sol before committing. Capability jumps between tiers vary by task type.
2. Profile reasoning token usage. Log total tokens (including reasoning) on a representative sample of 50-100 real requests before projecting monthly cost.
3. Default new projects to Terra, not Sol. Start mid-tier and escalate only where evals show a quality gap. Most structured tasks will not need Sol.
4. Re-test any Microsoft 365 Copilot prompts. If your team ships prompts into the M365 ecosystem, GPT-5.6 is now the runtime. Validate tone, format compliance, and tool-call behavior against the new model.
5. Watch the Models space for reasoning efficiency disclosures. OpenAI has not published per-task reasoning token benchmarks yet. Community profiling data will be the most reliable signal in the next two weeks.

The GPT-5.6 launch is a pricing and accessibility story as much as a capability one: frontier-tier AI just got cheaper, but only if your workload does not burn reasoning tokens faster than the rate card implies.