OpenAI announced GPT-5.5 on 23 April 2026, marking a significant model update with changes to reasoning capabilities and output behaviour [source]. The provider described the release as incorporating "enhanced multi-step reasoning" and "improved factual grounding," though specific technical details about architecture modifications were not disclosed.

The changelog indicates GPT-5.5 introduces a new default temperature setting of 0.7, down from the previous 1.0, which OpenAI states will reduce variability in responses. The model also implements what the provider calls "citation-aware generation," designed to surface source attribution more consistently in factual queries.

OpenAI reported that GPT-5.5 was trained on data through February 2026, extending the knowledge cutoff by approximately six months from GPT-5. The announcement noted "substantial improvements" in mathematical reasoning and code generation benchmarks, with the provider citing internal evaluations showing a 23% reduction in hallucination rates compared to the prior version.

The update affects all API endpoints using the `gpt-5` model identifier, which will now route to GPT-5.5 by default. Users requiring the previous version must explicitly specify `gpt-5-legacy` in API calls. OpenAI indicated this auto-upgrade behaviour aligns with its standard deployment practice for point releases.

The provider stated GPT-5.5 maintains the same context window length of 128,000 tokens and pricing structure as GPT-5. No changes to content filtering or moderation systems were mentioned in the announcement.

Developers relying on consistent model behaviour may need to validate outputs following this deployment, particularly in applications where temperature or reasoning patterns affect downstream logic.