Introduction

OpenAI released GPT-5.5 on April 23, 2026 as its smartest and most intuitive model to date, positioned as the new baseline for complex production workflows. The model is a strong fit for coding agents, tool-heavy assistants, grounded retrieval, product-spec-to-plan workflows, and customer-facing surfaces where execution quality matters. It matches GPT-5.4 per-token latency in real-world serving while reasoning at a higher level, and it uses meaningfully fewer tokens to complete the same Codex tasks.

This article distills OpenAI's official prompting guidance into a single reference. It covers what changed in GPT-5.5, how to write outcome-first prompts, how to shape personality and formatting, how to budget retrieval and tool use, how to drive coding agents, and a reusable prompt template at the end.

What Is New In GPT-5.5

GPT-5.5 should be treated as a new model family to tune for – not a drop-in replacement for GPT-5.2 or GPT-5.4. Begin migration with a fresh baseline rather than carrying over every instruction from an older stack. Start with the smallest prompt that preserves the product contract, then tune reasoning effort, verbosity, tool descriptions, and output format against representative examples.

The headline shifts compared to GPT-5.4 are:

• More Efficient Reasoning. GPT-5.5 reaches strong results with fewer reasoning tokens at the same effort setting. Re-evaluate low and medium effort before escalating; gains compound in tool-heavy, multi-step workflows.

• Outcome-First Task Execution. The model is better at working from a clear goal, preserving constraints, and turning product intent into concrete next steps. Step-by-step process instructions often hurt more than they help.

• Stronger, More Precise Tool Use. GPT-5.5 handles large tool surfaces, multi-step service workflows, and long-running agent tasks with more accurate tool selection and argument shaping.

• Polished But Direct Default Tone. Out of the box the model is efficient, direct, and task-oriented. Customer-facing or conversational products need explicit personality and warmth.

• Smarter Image Handling. When image_detail is unset or auto, GPT-5.5 preserves images without resizing up to 10,240,000 pixels or a 6,000-pixel dimension limit, improving computer-use performance.

• Reasoning Effort Defaults To medium. Treat medium as the balanced starting point. Reserve none for latency-critical, non-reasoning tasks. Use high or xhigh only when evals show measurable quality gains – higher effort is not automatically better.

Benchmark Snapshot

GPT-5.5 leads or ties the field on most agentic, coding, and research benchmarks reported in OpenAI's launch evaluations. Bold cells indicate the best score in each row.

Choosing Reasoning Effort

Reasoning effort is the single biggest dial on cost, latency, and quality. The default is medium; escalate only when evals justify it.

Outcome-First Prompts And Stopping Conditions

GPT-5.5 is strongest when the prompt defines the target outcome, success criteria, constraints, and available context, then lets the model pick the path. Describe the destination, not every step.

The loop you want the model to run looks like this – a tight cycle gated by a self-check, not by process scaffolding:

Prefer outcome-first phrasing:

Avoid stacking absolute rules. Older prompts overuse ALWAYS, NEVER, must, and only to keep weaker models on track. With GPT-5.5, reserve those words for true invariants such as safety rules, required output fields, or actions that must never happen. For judgment calls – when to search, ask, call a tool, or keep iterating – prefer decision rules.

Add explicit stopping conditions so the model knows when to commit:

Define missing-evidence behavior up front, so the model does not endlessly chase context:

Personality And Collaboration Style

For customer-facing assistants, support workflows, coaching products, and any other conversational surface, define both personality and collaboration style. Keep both short.

• Personality controls how the assistant sounds: tone, warmth, directness, formality, humor, empathy, polish.

• Collaboration Style controls how the assistant works: when it asks questions, when it assumes, how proactive it is, how much context it offers, when it checks itself, and how it handles risk.

Neither block should replace clear goals, success criteria, tool rules, or stopping conditions.

Example for a steady, task-focused assistant:

For more expressive products, add curiosity, humor, or a point of view explicitly – but keep the block compact. Personality shapes experience; it does not compensate for unclear goals or missing task instructions.

Improving Time To First Visible Token With Preambles

In streaming applications, users notice the delay before the first visible token. GPT-5.5 may spend time reasoning, planning, or preparing tool calls before emitting visible text. For longer or tool-heavy tasks, prompt the model to start with a short preamble: a brief visible update that acknowledges the request and states the first step.

The sequence below shows how a preamble fills the gap between request and first tool result:

A minimal pattern:

For coding agents that expose separate message phases:

Preambles also improve traceability: the model can state what it is about to check or do, then continue from that same assistant state after tool results arrive.

Grounding, Citations, And Retrieval Budgets

For grounded answers, citation behavior should be part of the prompt. Define what needs support, what counts as enough evidence, and how the model should behave when evidence is missing. Absence of evidence should not automatically become a factual "no."

Retrieval budgets act as stopping rules for search:

The same logic, as a decision tree the model can follow at each step:

Creative Drafting Guardrails

Drafting tasks – slides, launch copy, customer summaries, talk tracks, leadership blurbs, narrative framing – benefit from explicit separation of source-backed facts and creative wording.

Formatting And Verbosity

GPT-5.5 is highly steerable on output format. Set text.verbosity deliberately – the API default is medium, but low is often a better starting point for concise production responses. Heavier structure (headers, tables, nested bullets) should be reserved for cases where it improves comprehension or your UI needs a stable artifact.

A conservative default:

Add audience and length guidance when you need a specific register:

For editing, rewriting, or polishing customer-facing copy, tell the model what to preserve before asking it to improve style:

Tool Use, Phase Handling, And Responses State

GPT-5.5 works best in the Responses API. Use previous_response_id for multi-turn state. For stateless or Zero Data Retention flows, pass back the relevant output items each turn.

The phase field on assistant items distinguishes intermediate updates from final answers. When manually replaying assistant items, pass each phase back unchanged:

Key practices:

• Tool Descriptions Carry Most Of The Weight. Put what the tool does, when to use it, required inputs, side effects, retry safety, and common error modes inside the tool description itself. Reserve system instructions for cross-cutting policy.

• Prefer Hosted Tools Where They Fit. Web search, file search, code interpreter, image generation, and computer use reduce custom orchestration burden.

• Use Tool Search For Large Catalogs. Defer tool definitions and load only the relevant subset for the current turn.

• Phase Parameter. If your application manually replays assistant items instead of using previous_response_id, preserve phase values exactly: commentary for intermediate user-visible updates, final_answer for the completed answer. Do not add phase to user messages.

• Compaction. For long-running agents, compact intentionally: preserve completed actions, active assumptions, IDs, tool outcomes, unresolved blockers, and the next concrete goal.

• Prompt Caching. Keep stable content at the start of the request and dynamic user-specific context near the end. Use prompt_cache_key consistently and track usage.prompt_tokens_details.cached_tokens.

• Current Date. GPT-5.5 is aware of the current UTC date. Only add explicit date context for business timezones, policy-effective dates, or user-local references.

Coding Workflows

GPT-5.5 is OpenAI's strongest agentic coding model to date, reaching 82.7% on Terminal-Bench 2.0, 58.6% on SWE-Bench Pro, and outperforming GPT-5.4 on long-horizon internal coding evals while using fewer tokens. Codex is the canonical surface, but the prompting patterns transfer to other coding agents.

For coding agents, be explicit about reuse, subagent delegation, test expectations, acceptance criteria, and when to continue versus ask for help. Ask for concrete validation:

For implementation plans, make them traceable:

For visual artifacts, ask for inspection after rendering:

Migration Quickstart From GPT-5.2 Or GPT-5.4

The recommended migration order minimizes regressions: change the model, fix orchestration plumbing, then prune the prompt, then re-benchmark.

1. Update the model slug to gpt-5.5.

2. Use the Responses API for any reasoning, tool-calling, or multi-turn case.

3. Tune reasoning.effort: low for efficiency, medium as the balanced default, high for hard agentic tasks, xhigh for the hardest asynchronous work or evals.

4. Set text.verbosity to low if you previously relied on terse responses – on GPT-5.5, low is proportionally more concise than on GPT-5.4.

5. Verify that your application handles phase, preambles, and assistant-item replay correctly.

6. State the expected outcome and success criteria. Remove detailed step-by-step process guidance unless the exact path matters.

7. Move output schema definitions out of the prompt and into Structured Outputs.

8. Optimize for prompt caching: static parts first, dynamic parts last.

9. Drop the current date from the system prompt unless a non-UTC reference is required.

10. Benchmark against representative evals on accuracy, token consumption, and end-to-end latency.

OpenAI also ships an automated migration through Codex and the OpenAI Docs Skill:

Suggested Prompt Structure

Use this template as a starting point for complex prompts. Keep each section short. Add detail only where it changes behavior.

The seven sections fall into three layers – identity, contract, and exit conditions:

Closing Notes

GPT-5.5 rewards prompts that name the destination, hand the model a real budget of reasoning and retrieval, and trust it to choose the path. Many legacy prompt stacks shrink dramatically on this model – process scaffolding that was load-bearing on GPT-4.1 or even GPT-5.2 often becomes noise.

The practical playbook: start with the smallest prompt that preserves the product contract, write outcome-first goals with explicit stopping rules, push tool-specific guidance into tool descriptions, define personality and verbosity for any user-facing surface, and let evals tell you when to escalate reasoning effort. Treat every instruction as something the model will follow literally, and remove anything that is not earning its keep.