GPT-5.5: OpenAI’s new model that codes and reasons autonomously

GPT-5.5 is OpenAI's latest entry in the high-performance AI model segment. The company positions it as its most capable model to date, focused on complex tasks that require multi-step planning and execution: coding, office work, data analysis, and scientific research.

One of its most notable features is the ability to handle long-horizon tasks autonomously. Rather than waiting for step-by-step instructions, the model can take an ambiguous assignment, break it into subtasks, use tools, check intermediate results, and keep going until the task is done. OpenAI calls this "agentic" behavior, referring to the model's ability to act as an agent managing its own workflow.

In coding, GPT-5.5 scores 82.7% on Terminal-Bench 2.0, a benchmark for complex command-line workflows, compared to 75.1% for GPT-5.4, 69.4% for Claude Opus 4.7, and 68.5% for Gemini 3.1 Pro. On SWE-Bench Pro, which evaluates real-world GitHub issue resolution, it reaches 58.6% — below Claude Opus 4.7's 64.3% but above Gemini 3.1 Pro's 54.2%. Engineers with early access report that the model shows a stronger ability to understand the structure of complex software systems and anticipate problems without explicit prompting.

For office work, the model can autonomously operate computer interfaces, navigate between applications, and generate documents, spreadsheets, or presentations. OpenAI notes that over 85% of its employees already use Codex weekly. The company's finance team, for example, used the model to review more than 24,000 tax forms in less time than usual.

In scientific research, GPT-5.5 contributed to the discovery of a new mathematical proof about Ramsey numbers, later verified using the formal proof assistant Lean. In biological data analysis, it scores 80.5% on BixBench versus 74.0% for GPT-5.4, and on FrontierMath — an advanced mathematics benchmark — it outperforms both Claude Opus 4.7 (43.8%) and Gemini 3.1 Pro (36.9%) with 51.7% on tiers 1 to 3.

On the technical side, the model maintains latency comparable to GPT-5.4 despite being more capable, thanks to infrastructure optimizations built around NVIDIA GB200 and GB300 chips. OpenAI also states it uses fewer tokens to complete the same tasks, which reduces cost per use.

On safety, OpenAI has rated GPT-5.5's cybersecurity and biology capabilities as "high" under its risk evaluation framework, stopping short of the "critical" level. The model includes additional controls for potentially dangerous uses and offers expanded access to organizations working on critical infrastructure defense.

GPT-5.5 is available today for Plus, Pro, Business, and Enterprise plans on ChatGPT and Codex. GPT-5.5 Pro, designed for higher-accuracy work, is also live for Pro, Business, and Enterprise users. API access is coming soon, priced at $5 per million input tokens and $30 per million output tokens.

Key points

• Handles long, complex tasks autonomously without step-by-step instructions.
• Focused on coding, office work, and scientific research.
• Outperforms Claude Opus 4.7 and Gemini 3.1 Pro on Terminal-Bench 2.0 (82.7%).
• Falls below Claude but above Gemini on SWE-Bench Pro.
• More efficient than GPT-5.4: same latency, fewer tokens per task.
• Contributed to the discovery of a formally verified mathematical proof.
• OpenAI rates its cybersecurity and biology capabilities as "high".