OpenAI has released GPT-5.5, described as its first fully retrained base model since GPT-4.5. The update focuses heavily on accuracy and reliability, with OpenAI reporting a 52.5% reduction in hallucinations compared to the previous model, alongside a new lead on Terminal-Bench 2.0 at 82.7%.
What changed under the hood
Unlike incremental updates that fine-tune an existing base model, GPT-5.5 was retrained from scratch on a refreshed dataset and training process. OpenAI says this is what enabled the large drop in hallucination rate, rather than just patching specific failure cases.
Key improvements
- 52.5% fewer hallucinations: The model is significantly less likely to confidently state incorrect facts, especially on questions involving dates, statistics, and lesser-known topics.
- Terminal-Bench 2.0 leader: At 82.7%, GPT-5.5 currently leads this benchmark, which tests how well a model can complete real command-line and developer tasks.
- More consistent tone: Early users report more consistent response style across long conversations, with less drift in personality or formatting.
What this means for you
- If you use ChatGPT for research or fact-checking: You should see noticeably fewer confidently-wrong answers, but it’s still good practice to verify critical facts (dates, statistics, names) independently — “fewer” hallucinations doesn’t mean “zero.”
- If you use ChatGPT for coding: The Terminal-Bench improvements suggest better performance on tasks involving running commands, debugging build errors, and working with developer tooling.
- Free vs. Plus users: New base models are typically rolled out to paid tiers first, with free-tier access following on a delay or with usage limits — check your ChatGPT model picker to see if GPT-5.5 is available to you yet.
How it compares
GPT-5.5’s standout feature is reliability — the hallucination reduction is the largest jump reported by any of the three major labs in this release cycle. Gemini 3.1 Ultra instead emphasizes context window size (2M tokens) and native multimodal input, while Claude Opus 4.8 emphasizes long-running agentic and coding workflows with parallel subagents.
FAQ
How do I switch to GPT-5.5 in ChatGPT?
Open the model picker in the ChatGPT interface — if GPT-5.5 is available on your plan, it will appear as a selectable option, sometimes replacing the previous default model entirely.
Does GPT-5.5 cost more via the API?
Pricing for new flagship models is typically announced alongside release and may differ from the previous generation — check OpenAI’s pricing page for current rates.
Bottom line: GPT-5.5 is a ground-up retrain focused on trustworthiness — fewer made-up facts and stronger performance on real developer tasks.