OpenAI's GPT-5.5 Lands With Stronger Coding Numbers and Tighter Cyber Limits

OpenAI released GPT-5.5 on Thursday, April 23, calling it the company's "smartest and most intuitive" model yet. Two variants shipped, GPT-5.5 Thinking and GPT-5.5 Pro, both restricted to paying ChatGPT subscribers on the Plus, Pro, Business, and Enterprise tiers. API access opened a day later after OpenAI said it had added "different safeguards" specific to the new release.

OpenAI's headline numbers favor the new model on agentic coding. The company reports a score of 82.7% on Terminal-Bench 2.0 and 51.7% on FrontierMath categories 1 to 3, with both figures ahead of Anthropic's Claude Opus 4.7 and Google's Gemini 3.1 Pro on the same benchmarks. The model is also generally available through GitHub Copilot.

Independent testing tells a less tidy story. Tom's Guide reported that GPT-5.5 lost in all seven categories it ran head-to-head against Claude Opus 4.7, praising its speed but flagging persistent hallucinations. ZDNET was kinder, citing improvements in coding, conceptual clarity, and scientific reasoning over GPT-5. Benchmark wins, real-world losses: a familiar pattern in this generation of frontier releases.

Cyber safety as the headline feature

The more telling part of the launch is the safety packaging. OpenAI describes the release as carrying its "strongest set of safeguards to date," with new restrictions on high-risk cyber workflows, authenticated access controls for sensitive requests, and monitoring designed to catch repeated misuse patterns. The system card says the controls were tested with external experts.

The framing matters. OpenAI now treats offensive cyber capability as a release-gating concern rather than a feature to highlight, which echoes Anthropic's recent decision to hold back Project Glasswing on similar grounds. The frontier labs appear to be converging on a posture: ship the cyber capability, but build the guardrails first and tell the world about both.

OpenAI has not published parameter counts or context length, so direct architectural comparisons with DeepSeek's V4-Pro, which advertises a one-million-token window, remain guesswork. API pricing also has not been disclosed yet. Given that the cost of "good enough" inference has dropped roughly 50% since January, expect aggressive numbers when they land.

Coverage from CNBC, Wikipedia, and Help Net Security followed the launch in detail.