Why GPT‑5.6’s Real Breakthrough Is No Longer Hand‑Holding Prompts

Prompt simplification impact

OpenAI internal A/B testing compared the traditional verbose system prompt (role + steps + format + cautions) with a concise prompt that only states the desired result. The simplified prompts increased evaluation scores by 10‑15%, reduced token usage by 41‑66%, and lowered cost by 33‑67%.

ChatGPT Work

ChatGPT Work is a new feature that can stay attached to a project for hours, autonomously decomposing the task, retrieving data, writing content, and formatting output without step‑by‑step prompting.

Example prompt: "帮我做一份2026年上半年AI办公工具行业分析报告,包含市场规模、头部产品对比、趋势预测,输出Word格式". The model fetched data, created an outline, expanded each section, performed formatting, and produced a 3,000‑word Word document in roughly 20 minutes. The same workflow using manual stepwise prompting required at least 90 minutes.

Failure mode: when the model opens a browser to fetch data, an unstable network or VPN interruption can cause the job to stall.

Model tiers

GPT‑5.6 introduces three performance tiers:

Sol : flagship tier, best for complex reasoning and coding, highest cost.

Terra : balanced tier, suitable for everyday office work, priced at about half of Sol.

Luna : lightweight tier, optimal for batch processing and simple classification tasks.

Empirical observation shows Terra handles roughly 90 % of routine tasks with no noticeable quality loss compared to Sol. Sol is reserved for deep‑reasoning or heavy‑code scenarios, while Luna saves significant cost on high‑volume text‑classification jobs.

Quota sharing: ChatGPT Work and Codex now draw from the same quota pool, so heavy usage of Work reduces the quota available for Codex.

Prompt length comparison

verbose = """你是一个资深的行业分析师,拥有超过15年的科技行业研究经验。
请帮我写一份关于2026年上半年AI办公工具的行业分析报告。
要求:
1. 包含市场规模数据,需要有具体数字
2. 对比头部3-5个产品的功能和定价
3. 预测下半年趋势,给出至少3个方向
4. 语言要专业但通俗易懂
5. 每个章节不少于500字
6. 最后给出投资建议
7. 格式用Word文档
8. 注意数据来源要标注"""
concise = "帮我做一份2026年上半年AI办公工具行业分析报告,包含市场规模、头部产品对比、趋势预测,输出Word格式"
print(f"啰嗦版: {len(verbose)} 字")
print(f"精简版: {len(concise)} 字")
print(f"省了 {len(verbose) - len(concise)} 字,减少 {(len(verbose) - len(concise)) / len(verbose) * 100:.0f}%")

Result: 啰嗦版 192 字, 精简版 53 字, reduction 139 字 (72 %). The concise prompt produced a better output than the verbose version.

Another case: removing a long “Xiaohongshu viral writer” preamble and issuing a short request yielded more authentic Xiaohongshu‑style copy.

Practical actions

Delete boilerplate system instructions such as “You are an X expert” and state only the desired result.

Provide the full goal to ChatGPT Work instead of breaking it into multiple steps; let the model handle decomposition.

Select the appropriate model tier: Terra for daily tasks, Sol for heavy reasoning or coding, Luna for high‑volume batch jobs.