GPT-5.5 Launches — OpenAI's First Fully Retrained Model Since GPT-4.5
Axotron Take
This isn't an incremental update — it's a full retrain. GPT-5.5 leads on agentic workflows and terminal automation. The gap between labs just got real again. OpenAI needed this.
Full Notes
GPT-5.5 is OpenAI's first fully retrained base model since GPT-4.5. Every model in between was post-training updates — fine-tuning on top of the same base. This is different. A full retrain means new capabilities baked in at the foundation level, not patched on top.
The numbers back it up. 82.7% on Terminal-Bench 2.0, 78.7% on OSWorld, and strong agentic coding performance that senior engineers at NVIDIA described as feeling like "a limb amputated" when they lost access during testing.
Where it leads: multi-tool workflows, computer use, terminal automation, and research pipelines. Where it doesn't: raw coding depth (Claude Opus 4.7 still leads SWE-bench), cost efficiency (DeepSeek V4 is 34x cheaper), and long context above 256K.
The release came just days after DeepSeek V4 dropped open source at near-zero cost. OpenAI needed to remind the market that frontier performance still matters. GPT-5.5 does that. The race is not slowing down.