OpenAI details GPT-5.6's efficiency gains
OpenAI cut GPT-5.6 pricing and explained the inference and harness work behind it: Sol beats Claude Fable 5 on coding at less than half the cost.
OpenAI lowered GPT-5.6 pricing for its Terra and Luna models this week and published the engineering behind the cut. By OpenAI’s own benchmarks, flagship model GPT-5.6 Sol at max reasoning outperforms Claude Fable 5 on the Artificial Analysis Coding Agent Index at less than half the cost; Terra matches GPT-5.5’s intelligence benchmarks at half its price; and Luna, the fastest and cheapest model in the family, is priced 80% below Sol. OpenAI says the gains came from optimizing three layers at once: the models themselves (trained to do more work per token), inference (load balancing, speculative decoding, caching, and kernel-level optimization), and the agentic harness shared by Codex and ChatGPT Work, which cuts context bloat and repeated work. OpenAI also said GPT-5.6 Sol played “an instrumental role” in landing several of these optimizations autonomously. The company says it now serves more than 1 billion active users and over 2 million businesses.
What it means for operators
The concrete number to act on is the half-price-or-better claim at matched or better capability, not the framing. If your workloads are locked into GPT-5.5-era pricing assumptions, this is a prompt to re-benchmark against Terra and Luna directly rather than assuming last quarter’s cost-per-task numbers still hold; OpenAI is explicitly optimizing for exactly the kind of repeated agentic workloads, coding and long-running tasks, that most inflate token spend.