OpenAI officially unveiled GPT-5.6 in three tiers — Sol, Terra, and Luna — covering the full price spectrum from flagship to bulk. The biggest change from GPT-5.5: a 3× efficiency improvement, with comparable quality at 1/3 the token cost.
The tier breakdown:
- Sol (flagship): the strongest capability, on par with Claude Mythos on top benchmarks (MMLU-Pro 89.2, GPQA Diamond 78.4), priced at $15/$60 per million input/output tokens.
- Terra (mid): 90% of Sol's capability at 50% of the price ($5/$20), targeting medium-complexity tasks.
- Luna (bulk): optimized for high-volume, low-complexity tasks, priced at $0.5/$2, targeting search-augmented Q&A, classification, and extraction.
The technical path: GPT-5.6 uses a Mixture-of-Experts (MoE) architecture with 8B active parameters per token, drawing from a 200B total pool. The three tiers share the same MoE backbone but have different expert-selection strategies — Sol uses top-8 expert selection, Terra uses top-4, and Luna uses top-2. The result: same training, three products, very different cost profiles.
The industry impact: OpenAI is finally acknowledging that "capability per dollar" matters more than "absolute capability." The 3-tier strategy explicitly targets the Anthropic Claude series, the Google Gemini 3.x series, and the open-source Llama-4 / Qwen-3.7 series, with a clear "we can compete on price" signal.
For developers, the takeaway is that the era of "one model, one price" is ending. Tiered pricing will become the norm, and "model routing" — picking the right tier for each task — will be a new engineering discipline.