GPT-5.6 slashes prices: discounts up to 80%

OpenAI has reduced prices for its GPT-5.6 models. The most affordable model, Luna, now costs $0.2 per million input tokens and $1.2 per million output tokens, which is 80% cheaper than the previous price. The Terra model, designed for everyday tasks, has been discounted by 20%: now $2 for input tokens and $12 for output tokens. The flagship model, Sol, retains its original pricing ($5 and $30 per million tokens, respectively) but has gained a new Fast mode that increases speed up to 2.5 times compared to the standard mode, with the cost of such acceleration being double. OpenAI explains the price reduction by noting that Sol participated in optimizing its own production system: it rewrote GPU kernels, improved speculative decoding by 15%, and also participated in training monitoring, intervening when issues arose. As a result of these optimizations, service costs have dropped by 20%. The company emphasizes that the process is still human-in-the-loop: the model can write code and run experiments, but goals, result verification, and decisions on implementation remain with humans. The price cuts also affect Codex and ChatGPT Work products, and the auto-review model in these services has been replaced with GPT-5.6 Luna, which will reduce costs by approximately tenfold. These changes aim to lower the barrier to using agentic workflows, making them more accessible for frequent use.