OpenAI releases GPT-5.5, targeting coding and computer use

OpenAI shipped GPT-5.5 on April 23, less than two months after GPT-5.4. The gap between releases is shrinking: AI Release Tracker now logs 155 tracked frontier models since ChatGPT launched in November 2022, and the interval between major OpenAI versions has become a story in itself.

Greg Brockman, OpenAI's president, framed the release around autonomy rather than raw benchmark performance. "What is really special about this model is how much more it can do with less guidance," he told reporters Thursday. "It can look at an unclear problem and figure out just what needs to happen next." Practically, that means less hand-holding through each step of a task, which is the actual bottleneck in most real-world deployments.

CNBC reported the model handles a broad sweep of software-centric tasks: analyzing data, writing and debugging code, operating software interfaces, conducting online research, and producing documents and spreadsheets. Those workflows describe a large fraction of what engineering and applied science teams actually use AI assistants for today.

The safety picture

GPT-5.5 does not reach OpenAI's "Critical" cybersecurity risk threshold, which the company defines as enabling "unprecedented new pathways to severe harm." It does qualify under the "High" tier, meaning it can amplify existing harm pathways. OpenAI says it ran extensive third-party red-teaming on cyber and biological risks before release.

That distinction matters given the context. Anthropic's Claude Mythos Preview, announced earlier in April, triggered enough alarm about its ability to identify software vulnerabilities that Anthropic restricted its rollout. According to Price Per Token, OpenAI subsequently restricted access to GPT-5.5's most cyber-capable configurations as well, a notable shift from a company that had publicly criticized Anthropic's cautious handling of Mythos.

Mia Glaese, OpenAI's VP of research, addressed this directly during the Thursday briefing: "GPT-5.5 underwent extensive third-party safeguard testing and red teaming for cyber and bio risks, and we've been iterating on our cyber safeguards for months with increasingly cyber capable models." The word "iterating" is doing real work in that sentence. Safeguard development is framed as ongoing, not concluded.

The release picture

April 23 brought two simultaneous variants: GPT-5.5 and GPT-5.5 Pro. A lighter GPT-5.5 Instant followed on May 5, according to LLM Stats. All variants are rolling out to paid subscribers; OpenAI has not specified pricing beyond existing subscription tiers. The rapid cadence of variants suggests a differentiated product strategy rather than a single flagship model competing head-on with Google and Anthropic at every tier.

What it means for practitioners

The "less guidance required" claim is the technically interesting one. Agentic artificial intelligence systems have struggled in practice with ambiguous task decomposition. Models that require precise, step-by-step prompting do not scale to real engineering workflows where problems arrive underspecified. If GPT-5.5 genuinely advances this, it shifts where AI is useful: not just interactive sessions, but automated pipelines where a human cannot supervise every subtask.

Practitioners integrating GPT-5.5 into security-adjacent workflows or systems that interact with external software should treat the High cybersecurity risk classification as a design input, not a footnote. An artificial intelligence review of internal deployment policies before wide rollout is worth the time. The model documentation and OpenAI's usage policies will matter more in those contexts than headline benchmark numbers.

Release cadence at this level is not going to slow. The more consequential question is whether autonomous task performance on real, messy codebases improves fast enough to change how teams are structured. GPT-5.5's computer-use improvements point in that direction. Whether the gains hold outside controlled conditions is what the next months of production deployment will actually answer.

Frequently asked questions

What is GPT-5.5 and when was it released?
GPT-5.5 is OpenAI's latest AI model, released on April 23, 2026. It is optimized for coding, autonomous computer use, and research tasks, and is available to paid OpenAI subscribers. A lightweight variant, GPT-5.5 Instant, followed on May 5.

How does GPT-5.5's cybersecurity risk compare to Anthropic's Claude Mythos?
Both models prompted cybersecurity discussions in April 2026. Anthropic restricted Mythos' rollout over concerns about its ability to identify software vulnerabilities. GPT-5.5 does not reach the highest risk tier but qualifies at OpenAI's "High" level, and OpenAI later restricted access to its most capable cyber configurations as well.

What does OpenAI's "High" cybersecurity risk classification mean?
OpenAI classifies "High" risk as a model capable of amplifying existing pathways to severe harm without creating entirely new ones. It sits below the "Critical" threshold, which would trigger stricter release restrictions under OpenAI's own safety framework.

Is GPT-5.5 better at agentic tasks than previous models?
OpenAI claims it requires less user guidance for ambiguous tasks, which is the key bottleneck for agentic deployment. Whether that improvement transfers to uncontrolled production environments remains to be established through real-world use.