GPT-5.5 Long-Context Breakthrough: 74% on MRCR v2 at 512K-1M Tokens — 37-Point Jump Over GPT-5.4

OpenAI·medium signal

GPT-5.5's long-context performance represents its most dramatic improvement: 74.0% on MRCR v2 at 512K-1M token contexts, up from GPT-5.4's 36.6% — a 37-point jump. At 128K-256K tokens it scores 87.5% vs Claude's 59.2%. Combined with 72% fewer output tokens per task, this makes GPT-5.5 significantly cheaper for long-document agent workflows. For builders: if your agent pipeline processes large codebases or document collections, the cost-per-task improvement is substantial.