TL;DR
GPT‑5 cuts down factual mistakes in everyday use by combining smarter built‑in reasoning with training focused on honesty and safety. In OpenAI’s own tests, GPT‑5’s answers are ~45% less likely to contain a factual error than GPT‑4o, and when it “thinks” before replying, they’re ~80% less likely than OpenAI o3. There’s also new safety training (“safe‑completions”), stronger defenses against prompt injection, and specific tuning to reduce “sycophancy” (answers that blindly agree). OpenAI
What we mean by “hallucinations”
In AI, a hallucination is when the model asserts something that isn’t true—like inventing a citation or mixing up names and dates. GPT‑5 tackles this by (1) deciding when to answer fast vs. when to think longer, and (2) updating training and evaluation so the model is rewarded for being right (or admitting uncertainty), not for guessing. OpenAI+1
What changed under the hood
1) A router that chooses when to “think.”
GPT‑5 is a unified system with a fast default model, a deeper Thinking model, and a router that decides—based on your prompt and context—whether a quick response is enough or if the model should think longer. You can nudge it by saying “think hard about this,” and paid users can pick Thinking directly from the model picker. OpenAI+1
2) Measurably fewer factual errors.
OpenAI reports that, with web search enabled on real‑world prompts, GPT‑5’s answers are ~45% less likely to contain a factual error than GPT‑4o. When GPT‑5 uses its reasoning mode, the gap widens: ~80% fewer factual errors than OpenAI o3 on similar queries. On a claim‑by‑claim evaluation, the system card also shows GPT‑5’s main model has a 26% lower hallucination rate than GPT‑4o, while GPT‑5‑Thinking is 65% lower than o3; major‑error responses drop 44% (main) and 78% (thinking). OpenAI+1
3) Training that prefers safe, truthful output over brittle refusals.
GPT‑5 adds safe‑completions: instead of flipping between “answer” and “refuse,” the model is trained to give the most helpful allowed response, with clear, transparent refusals when needed. This reduces incentives to bluff and helps with tricky, dual‑use topics. OpenAI
4) Less “agreeing just to agree.”
To fight sycophancy, OpenAI post‑trained GPT‑5 to challenge shaky assumptions. In offline tests, the main model scored ~3× better than GPT‑4o on a sycophancy metric; early A/B tests showed 69% lower prevalence for free users and 75% for paid users compared to GPT‑4o. Fewer “you’re right!” answers when you aren’t. OpenAI+1
5) Tougher against prompt‑injection tricks.
On browsing, tool‑calling, and coding prompt‑injection tests, GPT‑5‑Thinking substantially outperforms OpenAI o3 (e.g., 0.99 vs 0.80 on tool‑calling injection score; higher is better), making it harder to derail with malicious instructions hidden in websites or tool outputs. OpenAI
Why it feels more accurate in normal chats
Most of the time you’ll just use Auto and GPT‑5 will decide how much effort to spend. If it switches to Thinking, you’ll see a lightweight reasoning view and can tap Get a quick answer if you’d rather it reply immediately. That balance—quick when it can, deeper when it should—is why routine questions feel faster while complex ones feel sturdier. OpenAI Help Center
How GPT‑5 reduces hallucinations (in plain English)
- Right level of effort, automatically. The router learns from real usage (what people pick, what they upvote, and measured correctness) and engages deeper reasoning only when it helps. Less guessing, more thinking where it counts. OpenAI
- Better grading & evals. OpenAI uses claim‑level grading (validated by people) and hard factuality benchmarks like LongFact and FActScore to drive improvements that show up in real queries—not just lab tests. OpenAI
- Honesty over bravado. Safe‑completions and anti‑sycophancy training reward admitting uncertainty, listing what’s missing, or offering a high‑level answer instead of over‑confident fiction. OpenAI+1
- Grounding when needed. In ChatGPT, you can turn on web search or link approved sources (e.g., Gmail/Calendar/Contacts via connectors). Pulling from the right place at the right time reduces the need to “make things up.” OpenAI Help Center+1
Get the most accurate answers: 7 copy‑paste tips
- Ask for depth when it matters.
“Think hard about this. Identify assumptions, unknowns, and trade‑offs before you answer. Cite sources where possible.” OpenAI - Require receipts.
“Use web search and give me 3 credible sources. Quote key lines in short snippets and link them.” - Constrain the task.
“Limit the answer to facts from [these docs/links]. If something isn’t in there, say you’re unsure.” - Make it show its work (briefly).
“List 3 alternative explanations and why you ruled them out.” - Prefer structured outputs.
“Return a table with claim, source, date checked, and confidence.” - Use the right mode.
For complex research or planning, switch to GPT‑5 Thinking (or tell Auto to “think hard”). For quick drafts, Fast/Auto is fine. OpenAI Help Center - Connect the evidence.
Enable Web search or relevant Connectors (Gmail, Calendar, Contacts) when you want the model to reference up‑to‑date or personal context safely. OpenAI Help Center
Where GPT‑5 still needs your judgment
Even with big gains, the model can still miss subtleties, over‑generalize, or cite sources that don’t fully support a claim. Treat it like a very capable analyst: give clear instructions, ask for sources, and make the final call—especially for health, legal, or financial decisions. (OpenAI stresses that ChatGPT does not replace medical professionals.) OpenAI
One‑minute FAQ
Does GPT‑5 really hallucinate less?
Yes. OpenAI’s public evaluations show large factuality improvements vs. GPT‑4o and o3, including ~45% and ~80% reductions in error rates on real‑world queries, plus big drops on dedicated factuality benchmarks. OpenAI
What changed to make that happen?
A smarter router, upgraded training (safe‑completions), explicit anti‑sycophancy tuning, and stronger defenses against prompt injection. OpenAI+2OpenAI+2
How do I force it to think more?
Type “think hard about this” or pick Thinking in the model picker. You can switch back to a quick answer mid‑stream. OpenAI+1
Sources
Connectors in ChatGPT (Help Center) (automatic Gmail/Calendar/Contacts connectors that can ground responses in your own data, if enabled).
Introducing GPT‑5 (model overview, router, factuality improvements and percentages, “think hard” hint). OpenAI
GPT‑5 System Card (safe‑completions, sycophancy metrics, prompt‑injection results, detailed hallucination analysis). OpenAI+3OpenAI+3OpenAI+3
GPT‑5 in ChatGPT (Help Center) (Auto/Thinking/Fast behavior, model picker, UI details). OpenAI Help Center
