OpenAI Launches GPT-5.5 as Strongest Agentic Coding Model

OpenAI released GPT-5.5 late Thursday, positioning the new model as its "strongest agentic coding model" to date and posting benchmark scores that extend its lead over rivals including Anthropic's Opus 4.7 and Google's Gemini Enterprise. Coverage ran across 23 and 24 April as the rollout reached ChatGPT users and the Codex coding assistant.

The Benchmark Headline

According to OpenAI's official release and reporting from CNBC, TechCrunch, MacRumors and TechStartups, GPT-5.5 scored 82.7 percent on Terminal-Bench 2.0 and 88.7 percent on SWE-bench, both standard tests of software-engineering capability. The model also posted 92.4 percent on MMLU, 73.1 percent on Expert-SWE, 84.9 percent on GDPval and 78.7 percent on OSWorld-Verified.

OpenAI reported a 60 percent reduction in hallucinations compared with the prior GPT-5.4.

Built for Agentic Work

The Tech Portal and TechStartups described GPT-5.5 as designed from the ground up as an "agentic model" that can work through complex tasks autonomously by switching between multiple tools, including computer use, research, coding and document or spreadsheet creation.

Rollout started immediately for ChatGPT Plus, Pro, Business and Enterprise users, plus the Codex coding assistant. GPT-5.5 Pro is limited to Pro, Business and Enterprise tiers. OpenAI said API availability will follow "shortly."

The Agentic Coding Race

The release lands as the AI industry reshapes around autonomous coding agents. It follows Anthropic's Opus 4.7 and Google's Gemini Enterprise launches earlier in the month, placing three major players in direct benchmark competition. Paired with the Intel/Terafab news out of Austin on the same day, the broader picture is a full-stack convergence from foundry to model that is redefining the AI infrastructure landscape.

For enterprise developers, the near-term question is whether the benchmark gains translate into sustained productivity in real codebases.