OpenAI introduces the GPT-5.

6 model family, consisting of three tiers: Sol (flagship, max reasoning outperforms Claude Fable 5 on the Artificial Analysis Coding Agent Index at less than half the cost), Terra (matches GPT-5.

5 on intelligence at half the price), and Luna (fastest, cheapest, 80% less than Sol).

The post emphasizes that efficiency gains come from optimizing across the entire stack—not just the model itself—focusing on inference and the agentic harness.

Inference optimizations include load balancing (GPT-5.

6 Sol in Codex analyzes production traffic to tune routing), kernel optimization (Sol autonomously rewrote production kernels in Triton and Gluon, reducing end-to-end serving costs by 20%), speculative decoding (Sol designed and ran experiments improving its draft model, increasing token-generation efficiency by more than 15%), and hyper-optimized KV cache configuration for specific workloads.

The agentic harness (a Rust orchestration layer) reduces repeated work by avoiding context bloat (deferred discovery of tools, capping tool output at 10,000 tokens), preserving exact prefixes for prompt caching (append-only history, deterministic tool order), and incremental transport.

The post notes these compounding wins allow OpenAI to deliver frontier intelligence and frontier efficiency simultaneously.

Sol’s autonomous optimization role makes the team optimistic about accelerating future gains.