GPT-5.5 Official Model Page: API and Tool Configuration

One-sentence takeaway

GPT-5.5 is suitable for connecting search, file, code interpreter, shell, computer use, and MCP tools through the Responses API; for deployment, pin a snapshot and first use medium reasoning effort as the cost/quality baseline.

Use cases

• Suitable tasks: Code agents, long-context retrieval, and research and knowledge work that requires web, file, or code tools.

• Unsuitable tasks: Tasks that only require image or audio output; the output modality listed on the model page is text.

• Applicable model versions: gpt-5.5 and the snapshot gpt-5.5-2026-04-23.

• Applicable clients, agents, or APIs: Responses API, Chat Completions, and Codex; prefer the Responses API for complex toolchains.

• Recommended reasoning tiers and parameters: none, low, medium, high, and xhigh all need to be checked against the current model page; GPT-5.5 defaults to medium. For production evaluation, fix the model snapshot, reasoning.effort, text.verbosity, and tool versions.

Ready-to-use content

Minimum configuration checklist (fill in your own deployment system):

model: gpt-5.5-2026-04-23
api: Responses
reasoning.effort: medium
input_modalities: text, image
output_modalities: text
context_window: 1,050,000 tokens
max_output_tokens: 128,000
knowledge_cutoff: 2025-12-01
tools: enable only the tools required by the task
acceptance: record output, tool calls, errors, latency, and token usage

API tool selection workflow:

1. Define the outcome and allowed side effects.
2. Enable only the required Responses tools: web search, file search,
   code interpreter, hosted shell, apply patch, computer use, MCP, or tool search.
3. Describe each tool's trigger, inputs, side effects, retry safety, and errors.
4. Start at reasoning.effort=medium and run a representative evaluation.
5. Keep the model snapshot and tool versions fixed while comparing effort levels.

Testing/workflow steps

1. Write the snapshot into the deployment configuration; do not use an unversioned alias as the sole reproducibility identifier.

2. Enable tools according to task permissions; do not expose the corresponding network, file, or computer-operation tools when they are not needed.

3. Document each tool's invocation conditions, parameters, visible side effects, error format, and retry safety.

4. Evaluate low, medium, and, when necessary, high separately with the same inputs; record success rate, latency, tokens, tool errors, and the amount of manual editing.

5. If the input approaches the million-token limit, separately measure retrieval hit rate, truncation, reasoning tokens, and long-context costs.

Original evidence and data

• The model page lists GPT-5.5's default reasoning as the highest tier/default medium API behavior and supports none, low, medium, high, and xhigh.

• The context window is 1,050,000, the maximum output is 128,000, and the knowledge cutoff date is 2025-12-01.

• The input modalities are text and image, and the output is text; the model page lists Responses and Chat Completions, and lists web search, file search, image generation, code interpreter, hosted shell, apply patch, computer use, MCP, and tool search among the tools available for Responses.

• Refer to the official pricing page for prices; the GPT-5.5 price corresponding to the model page is $5 per million input tokens and $30 per million output tokens. Long-context rules should be checked again before deployment.

Scope and limitations

• A tool being “available” does not mean the task will necessarily invoke it; trigger behavior should still be validated through tool descriptions and evaluation.

• The model page's knowledge cutoff date does not replace connected retrieval; time-sensitive facts should use controlled search or materials provided by the user.

• The context window is a limit, not a recommendation to fill it in one go; reasoning tokens, tool results, and multi-turn history also consume space.

• This configuration is not a security authorization checklist; permissions for computer use, shell, apply patch, and MCP must be minimized according to the environment.

Source excerpt or observation (compliance-limited short quote only)

The model page summarizes GPT-5.5's recommended uses as coding, tool-heavy agents, grounded assistants, and long-context retrieval; production environments still need to validate it with their own evals.