OpenAI Ships GPT-5.5 With Stronger Code and Computer Use

OpenAI released GPT-5.5 on Thursday, less than two months after GPT-5.4, extending a model cadence that has become the defining feature of the current competitive landscape. The new system comes in two variants, standard and Pro, and targets three capability areas: writing and debugging code, operating software autonomously, and conducting sustained online research.

Greg Brockman, OpenAI's president, framed the model's value in terms of autonomy rather than raw benchmark scores. Speaking at a press briefing, he pointed to the system's ability to interpret underspecified tasks and determine what needs to happen next without requiring detailed instructions. That framing positions GPT-5.5 less as a smarter chatbot and more as an early prototype of an artificial intelligence agent that can drive workflows with minimal human scaffolding.

According to CNBC, the model can analyze datasets, generate documents and spreadsheets, and operate software interfaces directly. The rollout starts with paid subscribers.

The cybersecurity picture

OpenAI's safety disclosure deserves close reading. GPT-5.5 does not reach the company's "Critical" cybersecurity tier, which it defines as enabling entirely new pathways to severe harm. It does meet the threshold for "High" risk, meaning it can amplify existing attack vectors in meaningful ways. Mia Glaese, OpenAI's VP of research, said the model went through extensive third-party red-teaming across cyber and biological risk categories, with safeguard iterations applied throughout the development cycle.

That classification lands in a charged context. The Hill reported that Anthropic restricted the rollout of its Claude Mythos Preview after the model demonstrated the ability to identify previously unknown vulnerabilities across major operating systems and web browsers, some dating back over two decades. Anthropic formed Project Glasswing to deploy those capabilities defensively, limiting access to vetted infrastructure firms. GPT-5.5 sits at a different point on that spectrum, but the High classification signals that OpenAI is navigating similar tensions between capability and risk.

Speed as strategy

The gap between GPT-5.4 and GPT-5.5 is itself a data point worth noting. LLM-Stats shows April 2026 as one of the busiest months on record for model releases: Alibaba's Qwen3.6-27B, Moonshot AI's Kimi K2.6, Anthropic's Claude Opus 4.7, and now two GPT-5.5 variants all shipped within days of each other. The pace is compressing the window that teams have to evaluate a model before the next generation arrives.

For applied teams, this creates a structural problem. The artificial intelligence review process that responsible deployment requires, covering safety, performance, and integration fit, now runs slower than the release cycle itself. GPT-5.5 Pro, listed as a separate release on Price Per Token on the same day, suggests OpenAI is also stacking capability tiers within a single generation, giving organizations more options but also more decisions to manage.

The competitive logic is transparent. Google and Anthropic are both pressing hard on autonomous coding and computer use, and Mythos Preview has drawn significant enterprise attention despite its restricted availability. Brockman's emphasis on doing more with less guidance is a direct bid for agentic workloads, where reduced prompt-engineering overhead translates into lower integration cost for the teams actually shipping these systems.

What practitioners should watch

Computer-use benchmarks tell only part of the story. A model that can drive software autonomously closes a meaningful gap between assistant and agent, but production performance depends on context that evaluation sets rarely capture. Teams evaluating GPT-5.5 for coding pipelines should focus on its behavior on underspecified inputs, precisely the area where OpenAI claims the sharpest improvement.

The High cybersecurity classification is also a practical compliance issue, not just an abstract risk label. Organizations in regulated industries or those with strict acceptable-use policies will need to work through what that rating means before deploying at scale. OpenAI's transparency here is more detailed than has been typical across the industry, but the underlying tradeoff remains real and cannot be discounted.

Open-weight models from Alibaba and Moonshot are matching proprietary systems on headline evaluations. The field is producing capable alternatives faster than ever. Whether that constrains OpenAI's pricing power or simply accelerates the whole ecosystem remains an open question, but the pressure to ship is clearly not easing for anyone.

Frequently asked questions

What is GPT-5.5 and what does it do?

GPT-5.5 is OpenAI's latest language model, released April 23, 2026. It targets autonomous computer use, code writing and debugging, online research, and document creation. It comes in standard and Pro variants and is available first to paid subscribers.

How does GPT-5.5 compare to Anthropic's Claude Mythos Preview?

Both models target coding and agentic tasks, but Anthropic restricted Mythos Preview to vetted infrastructure firms because of its cybersecurity capabilities. GPT-5.5 is available broadly to paid users, though it still carries a High cybersecurity risk classification under OpenAI's own framework.

What does a "High" cybersecurity risk classification mean for enterprise users?

OpenAI defines the High tier as models that can amplify existing attack vectors without opening entirely new ones. Teams in regulated industries should review that rating against their internal acceptable-use policies before deploying GPT-5.5 at scale.

Are open-source models competitive with GPT-5.5?

Open-weight models from Alibaba and Moonshot AI are scoring competitively on benchmark evaluations. GPT-5.5 may offer advantages in computer-use and agentic workflows, but direct comparisons depend heavily on the specific task and deployment context.