OpenAI Reasoning & coding
GPT-5.5
Turn a complicated brief into work you can review.
GPT-5.5 is OpenAI’s reasoning model for code, analysis and professional work. Give it the relevant source material, a clear outcome and explicit acceptance criteria. Use it to develop a patch, reconcile conflicting documents or turn an ambiguous requirement into an implementation plan. Connect execution tools when the task needs actions beyond a text response.
- Reasoning
- Text + image input
- Text output
Search, code execution and file creation need an integrated tool environment. Access to the model does not include a ChatGPT or Codex workspace.
AI Chat
Send a prompt and see the response here.
gpt-5.5Start a conversation with this model and ask follow-up questions.
USD /1M tokens
Pricing
| Group | Tier | Input | Output | Cache Read |
|---|---|---|---|---|
| default | input<272K | $5 | $30 | $0.5 |
| default | input>=272K | $10 | $45 | $1 |
| openai-0.2 | input<272K | $1 | $6 | $0.1 |
| openai-0.2 | input>=272K | $2 | $9 | $0.2 |
Public list prices are shown in USD. Final charges may vary by account group and usage tier.
Overview
What is GPT-5.5?
GPT-5.5 accepts text and images and produces text. OpenAI documents a 1,050,000-token context window and a maximum output of 128,000 tokens. The TokenHot catalog’s context value is shown above; treat provider limits and the limits of your chosen route as separate checks.
Start with work where several pieces of evidence need to agree: a bug report and a codebase, a specification and its implementation, or several versions of a business document. Decide how you will validate the result before sending the request. A model can propose a convincing explanation without having run a test or opened a source.
Make the change reviewable
Supply the failing behavior, relevant code and constraints. Ask for a minimal patch and a verification plan that another engineer can inspect.
Keep evidence beside the answer
Name your documents and require section references. Separate a documented fact from an inference or a missing piece of information.
Give tools a clear contract
Define the actions an agent may take, the output it must produce and when it must stop. Keep tool permissions in your application.
From brief to deliverable
Put GPT-5.5 to work
Engineering teams
Resolve a bug across a codebase
A failing integration often spans more than one function. Provide the reproduction, relevant modules and expected behavior so the model can connect the failure to a proposed change. Review the patch with the same tests and conventions you use for a human contribution.
- Tools you need
- Repository access, a shell and your test runner when you want changes executed. Pasted code is enough for a proposed patch.
- Keep in mind
- A proposed test is not a passing test. Require actual command output before accepting execution claims.
View a task brief
Use the attached bug report, relevant source files and test output to identify the cause of the failure. Propose the smallest maintainable patch that preserves the public interface. Include a regression check and explain why it covers the original failure. If a repository and test runner are available, apply the change and report the exact commands and results. Otherwise, return a patch for review and label every unexecuted check. Do not claim the issue is fixed without supporting evidence.
Product & operations teams
Reconcile conflicting project documents
Turn meeting notes, an older specification and a current change request into one decision memo. Assign each source a name and a date. Ask for disagreements and missing decisions explicitly, so a polished summary does not hide an unresolved requirement.
- Tools you need
- Text extraction or retrieval for source files; document software if you need a formatted export.
- Keep in mind
- The supplied sources define the evidence. Ask for missing information instead of allowing invented requirements.
View a task brief
Compare the supplied specification, meeting notes and change request. Produce a decision memo containing agreed requirements, conflicting requirements, unanswered questions and a proposed implementation sequence. Cite each factual requirement with its source name and section. Treat later dates as context, not automatic authority. Clearly label your recommendations. Acceptance requires that every requirement can be traced to supplied material and that no unresolved conflict is silently decided.
Frontend & quality teams
Turn a screenshot into an actionable issue list
Pair a screenshot with the expected flow and any available component code. Ask for concrete observations, likely implementation causes and a short list of checks. This is useful for preparing a handoff before a developer reproduces the issue in a browser.
- Tools you need
- Image input for visual analysis; a browser and assistive-technology checks for interactive verification.
- Keep in mind
- A screenshot cannot establish keyboard behavior, screen-reader output or a working submission flow.
View a task brief
Review the attached screenshot against the supplied UI requirements and component code. List the most consequential visible issues first. For each issue, describe the observed evidence, likely impact and a concrete implementation suggestion. Separate visible findings from hypotheses that require a browser. Provide a short keyboard, narrow-screen and form-submission checklist. Do not state that any interaction or accessibility check passed unless actual test evidence is supplied.
Capabilities in context
Key features
| Feature | What you get | Why it matters |
|---|---|---|
| Input and output | Text and image input; text output. | Analyze screenshots alongside text. Use a separate generation model or tool for images, audio or video. |
| Provider context | 1,050,000-token context; up to 128,000 output tokens. | Budget input and output together, and verify the route limit before sending a large request. |
| Reasoning control | OpenAI lists none, low, medium, high and xhigh; medium is the default. | Compare quality and latency on your own tasks. Confirm which parameter your TokenHot endpoint accepts. |
| API features | OpenAI documents streaming, function calling and structured outputs. | Your application still handles tool execution, schema validation and failed requests. |
| Version identity | OpenAI snapshot: gpt-5.5-2026-04-23. | TokenHot’s model ID is a platform alias. Confirm its upstream version if reproducibility requires a fixed snapshot. |
Provider specifications describe the OpenAI model. Confirm endpoint-specific options in the API docs before integration. OpenAI model reference
Published evaluations
Benchmarks, with context
Terminal-Bench 2.0
82.7%
Accuracy on command-line tasks involving tools and execution.
GDPval
84.9%
Wins or ties on the reported professional-work evaluation.
OSWorld-Verified
78.7%
Reported performance on computer-use tasks in an evaluation environment.
OpenAI-reported results from the April 23, 2026 release, using xhigh reasoning in a research environment. These are separate evaluations with different scoring rules, not TokenHot measurements or a guarantee for your workload. Evaluation results and conditions
Audience & fit
Where GPT-5.5 fits
Teams maintaining an existing integration
Keep a stable task set and compare accepted output, latency and cost before switching models. A version change deserves a fresh evaluation.
Builders of tool-based workflows
Use explicit tool schemas, bounded permissions and observable results. Put consequential actions behind the controls your application requires.
Analysts working from supplied evidence
Provide dated sources and a required answer structure. Ask the model to flag gaps and reconcile numbers instead of smoothing over disagreement.
Cost context
Plan the cost of a useful result
Count the complete request
Budget input, billable output and any applicable cache or tool charges. A short final answer can still involve additional reasoning. Track actual usage instead of estimating from visible answer length.
Check the long-input tier
The public TokenHot configuration places input at or above 272,000 tokens in a higher tier. OpenAI’s reference describes its boundary as above 272K. Confirm the platform’s condition at the exact boundary before a large job.
Measure an accepted result
Include retries, validation and tool calls in your comparison. Keep request content and success criteria consistent when evaluating a different model or account group.
Use the current pricing table for available rates and account groups. Model IDs and prices come from TokenHot’s catalog; provider pricing and subscriptions are separate.
API
Code examples
curl --location --request POST 'https://api.tokenhot.ai/v1/chat/completions' \
--header 'Authorization: Bearer <token>' \
--header 'Content-Type: application/json' \
--data-raw '{
"model": "gpt-5.5",
"messages": [
{
"role": "user",
"content": "What model are you"
}
]
}'FAQ
Frequently asked questions
What is GPT-5.5 best suited for?
GPT-5.5 is suited to the input, output, and capability types listed on this page. Test production workloads before choosing it for a critical system.
How is GPT-5.5 priced?
Pricing is shown in the pricing section above. Actual cost depends on usage volume, account group, and the request payload.
How can I call GPT-5.5?
Use the model ID shown above with one of the supported API protocols. Code examples are provided when usage data is available.
Is GPT-5.5 the same as GPT-5.5 Pro?
No. They are separate model variants. This page covers GPT-5.5. Use the model ID shown above for this route; do not assume Pro behavior or pricing applies.
Can GPT-5.5 generate images or videos directly?
The base model produces text. Image input lets it analyze visual material. Creating images or videos requires an appropriate generation model or integrated tool.
Does a GPT-5.5 API request include a browser or terminal?
No hosted workspace is implied. Your integration must supply the supported tools, permissions and execution environment needed by the task.
Why does the catalog context differ from OpenAI’s specification?
The card above reflects TokenHot’s current catalog; the feature table identifies the provider specification. A provider maximum is not a promise that every gateway route accepts that limit. Confirm the effective input and output budget before a large request.
How should I decide whether to keep GPT-5.5 or migrate?
Compare representative tasks with the same source material, tools and acceptance criteria. Record accuracy, required corrections, latency and total cost. Migrate when the new integration improves the result you actually need.
Get started
Build with GPT-5.5
API