GPT-5.5: ‘strongest agentic coding model ever’ failing spectacularly at its own game (LiveBench)

Oops!

"GPT‑5.5 is our strongest agentic coding model to date."

"The gains are especially strong in agentic coding."

"Instead of carefully managing every step, you can give GPT‑5.5 a messy, multi-part task and trust it to plan, use tools, check its work, navigate through ambiguity, and keep going."

These quotations sum up OpenAI's spin on 5.5. They created an entirely new subscription tier for it and made it the focus of Codex. Here, agentic coding isn’t just a feature but the selling point.

Well, looking at LiveBench’s independent agentic coding score, this is just a lot of hot air. The score for GPT-5.5 xHigh Effort is 56.67. Its predecessor, GPT-5.4, thrashes it at 70.00 on the same benchmark. Gemini 3.1 Pro, Claude 4.6 and others easily outperform it, too. In this highly relevant benchmark alone, it actually ranks 11th, just behind GPT-5.1 Codex.

While OpenAI were able to max Terminal-Bench (their benchmark) and SWE-Bench Pro, in a reliable test they didn’t design, select, or control, their main model falls drastically short compared both to its predecessor and the competition in the area it was meant to excel in.

Is this as damning as it looks? What's your experience actually using 5.5 for agentic coding?