OpenAI made GPT-5.5 Instant available to all users on July 22, including free-tier accounts. According to the official announcement, the model reduces hallucinated claims by 52.5% in high-risk scenarios such as healthcare, legal, and finance. In conversations flagged by users as factually incorrect, the rate of false statements dropped by 37.3%.
Mathematical reasoning also took a leap. The model scored 81.2 points on the AIME 2025 test, up from the previous generation's 65.4 points. AIME serves as a stress test for chain-of-thought integrity, meaning the model can now handle multi-step logic more reliably.
The rollout covers all users since May 5, including free accounts. However, personalized memory features—such as referencing past conversations, uploading files, or connecting Gmail—remain locked to Plus and Pro subscribers and are currently web-only.
How OpenAI Says Prompts Should Be Written
Just days before the GPT-5.5 Instant launch, OpenAI also published an official recommended prompt structure guide. The company argues that most people write prompts with fundamentally flawed logic.
The guide suggests seven sequential blocks: Role, Personality, Goal, Success criteria, Constraints, Output, and Stop rules.
The first key shift is "result-first" thinking. Old prompts tell the model step-by-step: do A, then B, then output C. The new approach defines the endpoint first—spell out the success criteria and let the model choose its route. OpenAI explicitly recommends deleting procedural steps and replacing them with outcome descriptions.
Another critical element is reasoning effort, a setting that controls how deeply the model thinks before answering. Higher effort means higher cost. The official advice is to use low or medium for most production scenarios; only multi-step reasoning, formatted output, or data extraction warrants the high setting—anything else is just burning money.
Additional tips include:
- Write stop rules to define when the model should cease, e.g., stop after finding the first qualifying result.
- Set retrieval budgets to cap the number of searches the model performs, preventing infinite loops.
- For draft tasks, define what you don't want—negative constraints are often easier for the model to parse than positive descriptions.
The simultaneous release of GPT-5.5 Instant and the prompt guide reflects OpenAI's dual focus on model capability and user interaction efficiency.