Stop being so demanding: OpenAI’s new GPT-5.6 guidelines change everything

OpenAI published a New claim guide For GPT-5.6 Sol, it is Newly released flagship modeland the main message will ring true to anyone who has spent the last year writing a multi-page system that demands: Stop writing so much. The basic idea is to claim the result first. Find good form, set your stopping conditions, and get out of the way.

Detailed how-to instructions, repetitive style rules, examples that don’t change behavior – all of this is now considered noise.

OpenAI backs this up with numbers: in internal crypto proxy tests, the leaner system claimed to improve evaluation scores by approximately 10-15% while reducing total tokens by 41-66% and costs by 33-67%.

GPT-5 vs. GPT-5.6: What’s Really Changed?

the GPT-5 claim guidewhich was published at launch in August 2025, was all about adding scaffolding. You’ve got persistence XML blocks that tell the model to keep running until the problem is resolved, detailed context-gathering templates that specify exactly how to balance searches and when to escalate them, and introductory scripts for the tool that narrate each step aloud.

The philosophy was to calibrate enthusiasm, by constructing clear lines that defined when to exert more effort or stop.

GPT-5.6 often doesn’t need those rails. The new guide tells you what’s going on: repetitive rules, pattern instructions that don’t change behavior, examples that do nothing, and processing steps that the model already handles reliably. So, basically, this block with parallel search batches and early stopping criteria that was used to help is now scaffolding for the model to analyze, not scaffolding to help it.

What it actually keeps is simpler: the user-visible score, success criteria, stopping conditions, and hard constraints. The evidence template for a good claim begins with “solving the customer’s problem from start to finish”—then specifies exactly what form to take, what actions must be completed before responding, and what to do when the required evidence is missing. Don’t “be inclusive”. Not “continue.” Just: here is the destination.

Risk calculations have also changed. The guide warns that GPT-5.6 follows spot contracts too closely, and that “conflicting rules can lead to greater instability than loss of detail.”

The previous model selected one instruction when a conflict occurred. GPT-5.6 burns logical codes trying to reconcile the two, is slower, more expensive, and often faulty. If your system prompt has nested rules — and most production prompts do — this is the thing to fix first.

OpenAI also strongly advises against using the old trick of resorting to absolute statements like “always do this” or “never do that” to steer AI behavior in a particular direction.

Two tangible additions complete the difference. The first is the text.verbosity parameter: Because GPT-5.6 is more concise by default than GPT-5.5, the old “be brief” instructions now overcorrect and make responses too short. Set the global default via the parameter, and then override each task at the prompt. The second is a section on programmatic tool calling – for limited workflows where code handles filtering, aggregation, or aggregation of large intermediate outputs and returns a compressed result, offloading that work from model rule entirely.

But does it work?

We used the evidence to improve our claim for TYPE OR DIE, a first-person survival horror game that we are building to benchmark the model’s programming capabilities. The result was more polished: GPT-5.6 Sol handled the auto-aim logic more efficiently than previous outings, the visuals were more cohesive, and the overall feel of the game was sharper.

It took more time to build. The model did not jump directly into programming, but mapped out the entire problem first, planning each system before writing a line. This is a guide that works as intended. Determine the destination; The model chooses the path.

The new prompt is available on our Github so you can check it out check it out.

You can play the original GPT 5.6 game by clicking This link.

The game created under the newer wave, is Available here.

If you want to go further, or are too lazy to memorize all these new instructions, you can create your own custom GPT and feed it with the full directory as its own knowledge base. Configure it to parse any prompt you give it, understand the underlying logic, and rewrite it in GPT-5.6 style. You end up using agile engineering to design better claims.

Urgency. You’re welcome.