OpenAI says GPT-5.6 works better with shorter prompts

OpenAI is telling developers to write less when using GPT-5.6 Sol, its newly released flagship model. For investors watching AI costs, the shift matters because shorter prompts can reduce token usage, which is one of the direct cost drivers for AI apps.

In a dedicated prompting guide for GPT-5.6 Sol, OpenAI says users should move toward outcome-first prompting: define the result, spell out what counts as done, set clear stopping conditions, and avoid loading the model with extra process instructions. A token is a unit of text processed by an AI model, and more tokens generally mean more compute and higher API bills.

OpenAI said internal coding-agent tests found that leaner system prompts lifted evaluation scores by about 10% to 15%. The company also said those shorter prompts cut total tokens by 41% to 66% and reduced costs by 33% to 67% in those tests.

What changed from GPT-5

OpenAI’s GPT-5 prompting guide, published at that model’s launch in August 2025, leaned on more structure. According to OpenAI’s materials, that earlier approach included XML-style persistence blocks, detailed templates for gathering context, instructions for parallel searches, escalation steps and tool preambles that described what the model was doing.

The GPT-5.6 guidance points in the other direction. OpenAI says repeated rules, style notes that do not affect behavior, examples that do not change the output, and process steps the model can already handle should be removed.

The practical version is simpler: tell the model the user-visible outcome, the criteria for success, when to stop, and any hard limits it must respect. OpenAI’s guide frames a strong prompt around resolving a customer issue from start to finish, then defining what completion requires and what to do if required evidence is missing.

Why prompt clutter can cost more

OpenAI says GPT-5.6 follows prompt contracts closely. That means a messy or contradictory system prompt can create problems because the model may spend extra reasoning tokens trying to satisfy instructions that do not fit together.

Reasoning tokens are the model’s internal work before producing an answer. If an AI model burns more of them reconciling overlapping rules, the result can be slower and more expensive, and OpenAI says it can also hurt reliability.

The guide’s warning is direct: conflicting rules can create more instability than missing detail. For companies building AI agents, customer-service tools or coding assistants, that puts prompt maintenance closer to cost control and product quality than cosmetic writing.

New tools in the guide

OpenAI’s GPT-5.6 guide also adds a first dedicated section on Programmatic Tool Calling, a feature area not included in the GPT-5 playbook, according to the company’s materials. Tool calling is when a model invokes software functions, such as search, code execution or external services, instead of only generating text.

The guide also highlights the text.verbosity API parameter, which was absent from the GPT-5 prompting guide. That parameter gives developers a direct way to influence how much text the model returns, rather than relying only on written style instructions inside the prompt.

For users and builders, OpenAI’s message is that GPT-5.6 needs clearer goals and fewer rails. The company’s own testing suggests that cutting unnecessary instructions can improve results while lowering the amount of text the model has to process.

This story draws on original reporting from Decrypt.