GPT-5.5
OpenAI · openai family · Official docs
GPT-5.5 is now the default OpenAI flagship target in Refrase. The adapter keeps the same OpenAI family rules as GPT-4o, but prompts should be more explicit about hierarchy, constraints, and success criteria because modern GPT models are more literal and more capable at long-context and agentic workflows. Refrase does not claim measured lift for GPT-5.5 yet; it is documentation-supported and queued for the next benchmark pass.
Specifications
Strengths
Key capabilities
- ✓OpenAI's latest flagship model for advanced coding, agentic task performance, and long-context understanding (source: OpenAI model docs, GPT-5.5)
- ✓Supports text and image inputs with text output on the Responses and Chat Completions APIs (source: OpenAI model docs, GPT-5.5)
- ✓256K context window and 128K max output tokens for long-form work (source: OpenAI model docs, GPT-5.5)
Known limitations
- ⚠Use GPT-5.5 Pro only when maximum intelligence is required; GPT-5.5 is the default flagship for most advanced work (source: OpenAI model docs, GPT-5.5)
How to prompt GPT-5.5
Preferred instruction format
OpenAI GPT models use role-based messages. Put durable behavior and hierarchy in system/developer instructions, then keep task-specific data in user content. For Responses API workflows, prefer first-class tools and structured output settings over embedding tool schemas or JSON schemas in plain prompt text.
Recommended practices
- State the objective, constraints, and output format explicitly; newer GPT models follow literal instructions more closely than legacy GPT-4o-style prompts.
- Use Markdown headings or XML tags to delimit role, context, task, constraints, examples, and output format.
- For agentic tasks, include persistence, tool-use honesty, and planning instructions so the model keeps working, avoids guessing, and plans before calling tools.
- Use native structured outputs and tool definitions through the API instead of pasting schemas into the prompt when possible.
Anti-patterns to avoid
- Do not rely on GPT-4o-era vague prompts like 'make this better'; define success criteria and output shape.
- Do not paste tool schemas into the natural-language prompt when first-class API tools are available.
- Do not overuse all-caps urgency or coercive language; prefer clear priority order and measurable constraints.