OpenAI Ships GPT-5.6: Shorter Prompts, Lower Costs, and a Warning About Brevity

OpenAI's latest model works better with minimal prompts, reduces token usage by up to 66%, and asks developers to rethink how they instruct AI.

OpenAI has released GPT-5.6, and the developer guidance that ships with it points to a quiet but significant shift in how we interact with large language models.

The most striking claim in the developer migration guide: internal evaluations found that replacing long, explicit system prompts with minimal prompts improved benchmark scores by 10 to 15 percent, while reducing total tokens by 41 to 66 percent and cutting costs by 33 to 67 percent. Less instruction produced measurably better results.

GPT-5.6 also preserves original image dimensions when processing uploaded images, rather than resizing them to a fixed patch budget, improving detail retention for vision tasks. The model infers user intent more accurately without needing every step spelled out, though OpenAI cautions developers to still state constraints and success criteria explicitly.

One eyebrow-raising note: GPT-5.6 is more sensitive to generic brevity instructions like "Be concise." The model appears tuned for more calibrated verbosity out of the box, and piling on terseness directives may produce unexpectedly clipped output. Separately, prompting for friendliness or empathy no longer meaningfully changes behavior.

A deployment safety evaluation accompanies the release, covering risks and mitigations under the voluntary safety framework OpenAI adopted after GPT-4.

On Hacker News, the release drew over 1,170 points and more than 800 comments. Commenters expressed enthusiasm for the cost savings but also wariness that prompt engineering habits built over two years may need overnight revision. One developer captured the sentiment: "I'd be stunned if GPT-5.6 was that concise by default. What if you were expecting GPT to be as wordy as it usually is?