GPT-5.5: The First Ground-Up Rebuild Since GPT-4
Released April 2026, GPT-5.5 is the first fully retrained OpenAI base model in two years. Here's what changed, what the benchmarks show, and how it compares to Claude.
June 19, 2026
GPT-5.5 landed on April 23, 2026 — the same day as DeepSeek V4, which made for an unusually competitive news cycle. OpenAI positioned it explicitly as the first fully retrained base model since GPT-4: not a fine-tune, not an incremental update, but a ground-up rebuild. By May 5 it had replaced every previous model as ChatGPT’s default.
What's New Under the Hood
- Context window: 1,050,000 tokens (1.05M) — slightly wider than Claude’s 1M
- Multimodal: unified architecture for text, images, audio, and video in a single model
- Pricing: $5.00/$30.00 per million input/output tokens
- Became ChatGPT’s default on May 5, 2026 (internal codename: Spud)
Benchmarks: Where GPT-5.5 Leads and Where It Doesn't
GPT-5.5 leads convincingly on long-context retrieval (MRCR v2: 74.0% vs Claude Opus 4.7’s 32.2%), general reasoning (ARC-AGI-2: 85.0% vs 75.8%), and agentic task completion (Terminal-Bench 2.0: 82.7% vs 69.4%). Claude leads on hard coding (SWE-Bench Pro: 64.3% vs 58.6%) and knowledge-intensive reasoning (Humanity’s Last Exam: 46.9% vs 41.4%). The two models are genuinely competitive — which is the best possible outcome for developers.
Bottom line