GPT-5.5

GPT-5.5 is an OpenAI frontier model built for complex, execution-heavy work: coding, online research, document creation, spreadsheet work, and multi-step tasks that span tools. OpenAI positions it as a model that understands goals earlier, uses tools more effectively, checks its own work, and carries tasks further before handing control back to the user.

Why it matters

GPT-5.5 represents a shift from models that merely answer well to models that can keep working through a job. That makes it especially relevant for agentic systems, coding agents, and professional workflows where persistence, context handling, and tool use matter as much as raw reasoning quality.

Official benchmark profile

At launch, OpenAI reported several strong benchmark results for GPT-5.5:

• Terminal-Bench 2.0: 82.7%, which OpenAI presented as state-of-the-art for complex terminal workflows
• SWE-Bench Pro: 58.6% on real-world GitHub issue resolution
• GDPval: 84.9% on professional knowledge work across occupations
• OSWorld-Verified: 78.7% for computer-use tasks in real software environments
• BrowseComp: 84.4% on tool-assisted browsing and web research

OpenAI's public comparison table also showed GPT-5.5 ahead of GPT-5.4 on these evaluations while using fewer tokens on coding tasks.

Practical tradeoffs

At launch, OpenAI said GPT-5.5 would come to the API with a 1 million token context window and pricing of $5 per million input tokens and $30 per million output tokens. That places it in the premium tier: powerful enough for difficult work, but expensive enough that teams usually reserve it for high-value steps like planning, escalation, code review, or long-horizon agent tasks.

In practice, GPT-5.5 is the kind of model you use when a cheaper model is likely to fail quietly: subtle bugs, fragile tool use, or incomplete multi-step execution. Its value is less about sounding smarter in chat and more about finishing hard work with fewer interventions.