OpenAI has released GPT-5.5 Instant, an update to the default ChatGPT model that reduces hallucination rates by 40% while introducing per-user calibration controls for context retention, tone consistency, and domain-specific guardrails. The upgrade is rolled into ChatGPT's default endpoint and does not require a separate subscription or opt-in.
What it does
GPT-5.5 Instant is a direct replacement for the previous default model. The headline improvement is a 40% reduction in hallucination rates — meaning the model fabricates facts, citations, or logic less often than its predecessor. OpenAI achieved this through a combination of improved training data filtering and reinforcement learning from human feedback (RLHF) adjustments.
More significant for power users is the new preference tuning system. For the first time in a frontier model, users can adjust:
- Context retention: how far back the model remembers conversation history.
- Tone consistency: whether the model stays formal, casual, or neutral.
- Domain-specific guardrails: custom restrictions on topics, sources, or output formats.
These controls are granular and per-user, not global. They do not require API access — they are available directly in the ChatGPT interface for all users on the default model.
Tradeoffs
OpenAI has not published latency benchmarks for GPT-5.5 Instant, but early reports suggest response times are comparable to GPT-4o. The preference tuning does not add noticeable delay because it is applied at inference time via lightweight parameter adjustments rather than full model retraining.
The 40% hallucination reduction is a claimed figure based on internal evaluations. Independent benchmarks have not yet been published. Users should still verify critical outputs, especially for factual claims, code, or medical/legal advice.
When to use it
GPT-5.5 Instant is now the default model in ChatGPT. No action is required to switch. Users who want the old behavior can revert to GPT-4o in the model picker, but OpenAI recommends staying on the default for most tasks.
The preference tuning is most useful for:
- Customer-facing chatbots that need consistent brand tone.
- Research assistants that must stay within specific source domains.
- Long-running conversations where context retention matters.
Bottom line
GPT-5.5 Instant is a meaningful incremental upgrade that addresses two of the biggest complaints about large language models: hallucination and lack of personalization. The 40% hallucination reduction is a strong claim that will need independent verification, but the preference tuning feature alone makes this worth updating for anyone who uses ChatGPT regularly.