OpenAI published a guide to independent tips for the newly released flagship model GPT-5.6 Sol, the core change being to weaken the long system tips that were common in the past. The new guide recommends that the phrase should first identify the objectives, the completion criteria and the conditions of cessation, rather than the description of the stacking process, the repetition of rules and style requirements.

Internal test gives data

The internal test data given by OpenAI show that in the coding proxy scenario, a streamlined system hint increases the rating by about 10 to 15 per cent. At the same time, total token consumption decreased by 41 to 66 per cent, and costs decreased by 33 to 67 per cent.

Move from adding process to writing results

OpenAI has previously placed greater emphasis on establishing clear processes for the model, including ongoing implementation blocks, context collection templates, and a description of the steps before the tool is called. The objective of such an approach is to allow the model to move along the established path in complex missions.

In the new GPT-5.6 guide, however, such “scaffolds” are significantly weakened. OpenAI believes that the model has been able to stabilize many intermediate steps and continues to include a large number of process descriptions, which would instead increase the analysis burden.

  • First, write the final results for the user.
  • Clarification of the criteria for completion of the mandate and conditions for cessation
  • Only the mandatory restrictions that must be maintained

Conflict rules are highlighted.

OpenAI also warns in the guide that GPT-5.6 will comply more strictly with the instructions. If there are overlapping or conflicting rules in the same system hint, the model consumes more reasoning token to coordinate these requirements, and the result may be slower, more costly and more prone to error.

This means that the usual “never do it” or “never do it” absolute formulations in the past are no longer recommended. The rule of conflict should be deleted and given a higher priority for optimization than continued restrictions.

Add two new usage changes

The guide also includes two elements not previously included in the GPT-5 guide. One is... Text.verbosity argumentsI don't know. OpenAI indicates that the default GPT-5.6 output is simpler than GPT-5.5 and that it may be too short if "short answers" continue to be requested repeatedly in the hint.

The second is the section “Procedure tools call”. This part applies to a clear flow of work at the boundary, such as the completion of screening, batch processing or aggregation by code, and then the submission of compressed results to the model for processing in order to reduce the burden on the model to determine its own intermediate steps.

Decrypt mentioned in the text that its team had used the new guide to optimize a game item tip to test model coding capabilities. According to its description, GPT-5.6 Sol performed better in automatic targeting logic, image consistency and overall completion, but earlier planning took longer and the model would be decomposed before the coding phase.