GPT-5.5 Is Here. It's Fast, It's Expensive, and It Changes How You Prompt.

OpenAI shipped GPT-5.5 on April 23, six weeks after 5.4. The capability jump is real: 82.7% on Terminal-Bench 2.0 vs Claude Opus 4.7's 69.4%. The Pro tier nearly doubles Opus 4.7 on FrontierMath Tier 4 at 39.6% vs 22.9%. It uses 40% fewer tokens on Codex tasks while matching 5.4's latency. Native desktop navigation baked in. Clicking buttons, typing text, multi-step workflows out of the box. OpenAI

The price doubles too. Standard API hits $5/$30 per million input/output tokens. Pro: $30/$180. OpenAI is betting the capability gap justifies the premium. Meanwhile, DeepSeek V4-Pro launched the same week at $1.74/$3.48 per million tokens. That's a 6-8x cost difference while V4-Pro posts 80.6% on SWE-bench, roughly matching GPT-5.5's comparable score. VentureBeat

Here's what caught me off guard. OpenAI's official prompting guide for 5.5 tells you to throw away everything you learned about prompt engineering for GPT-4 and GPT-5. Define target outcomes and constraints. Let the model pick the path. New text.verbosity controls at the API level steer response length programmatically. Simon Willison tested it with his pelicans benchmark and found the default lagged behind 5.4, but the xhigh reasoning effort level improved things dramatically at the cost of way more tokens.

The market is bifurcating. Frontier labs raise prices banking on capability. Open-weight models close the gap at commodity rates. For builders running production agent pipelines, model routing isn't optional anymore. Use GPT-5.5 or Opus 4.7 for hard reasoning. Route everything else to V4-Flash at $0.14/M input tokens. If you're not routing by task difficulty, you're burning money.