What is GPT-5.3-Codex?
GPT-5.3-Codex is a coding-specialized reasoning model provided by OpenAI. It is designed for software engineering and other computer-based professional tasks in which the model must do more than generate an isolated code snippet. In a suitable agent environment, it can examine a repository, reason about a requested change, call tools, run tests, investigate failures, and refine its work over multiple steps.
OpenAI positions GPT-5.3-Codex as a combination of the coding capabilities associated with GPT-5.2-Codex and the reasoning and professional knowledge capabilities of GPT-5.2. The model is primarily intended for Codex and comparable tool-enabled workflows rather than simple, one-shot text completion.
Its model identifier for API use is gpt-5.3-codex. The model was released on February 5, 2026 and is documented as current and available through OpenAI API and Codex surfaces.
Where GPT-5.3-Codex fits
GPT-5.3-Codex sits in OpenAI's current catalog as a coding-focused member of the GPT-5.3 family. Its design emphasizes long-running, agentic work: tasks in which a model can maintain context, make intermediate decisions, interact with development tools, and communicate with a user while the task is still in progress.
This positioning makes it different from a model selected mainly for low-cost classification, short answers, or high-volume text transformation. GPT-5.3-Codex is most useful when the work contains several dependent steps and the quality of the implementation matters more than minimizing every unit of latency or token cost.
Core capabilities and limits
| Specification | GPT-5.3-Codex |
|---|---|
| Context window | 400,000 tokens |
| Maximum output | 128,000 tokens |
| Input | Text and images |
| Output | Text only |
| Reasoning effort | Low, medium, high, and xhigh |
| Knowledge cutoff | August 31, 2025 |
| Fine-tuning | Not supported |
The 400,000-token context window allows a long task to include substantial code, documentation, test output, and conversation history. Context capacity is not the same as guaranteed correctness: a large repository still needs to be explored and managed carefully, and the model can produce inaccurate or incomplete work.
The maximum output limit is 128,000 tokens. This is an upper limit rather than a requirement or a promise that every request will produce a response of that size. In normal development, the model may use shorter responses while making tool calls or presenting a proposed change.
Reasoning for long-running work
GPT-5.3-Codex supports four reasoning-effort settings: low, medium, high, and xhigh. These settings provide a practical quality-versus-latency trade-off. Lower effort can be appropriate for interactive coding assistance or relatively clear changes. High and xhigh are intended for difficult engineering tasks that require more extensive analysis.
OpenAI's guidance recommends medium reasoning effort as a general interactive setting. High or xhigh may be more suitable when the task involves complex debugging, architectural changes, broad repository modifications, or difficult tool-driven execution. The trade-off is that deeper reasoning can increase response time and token consumption.
The model also supports compaction for extended sessions. Compaction is intended to preserve useful task state as a long interaction grows, helping an agent continue working without treating the entire uncompressed conversation as its only source of context.
Coding and agentic workflows
GPT-5.3-Codex is intended for autonomous or semi-autonomous software development. Examples supported by its positioning include implementing a feature across multiple files, tracing and repairing a failing test, reviewing a pull request, building a web application, analyzing data, preparing technical documentation, and operating a development environment through tools.
A typical workflow might begin with a request to update an existing application. The model can inspect relevant files, identify dependencies, propose an implementation, make edits through an available tool, run tests, interpret errors, and revise the changes. The model's value in this setting comes from coordinating those steps rather than merely predicting a block of source code.
These workflows require an appropriate harness or product integration. The model does not independently gain unrestricted access to a computer, repository, network, or deployment system simply because it supports tool use. The surrounding application must provide the tools and determine what permissions, files, commands, and external services are available.
Tools and API support
OpenAI documents GPT-5.3-Codex for use with the Responses API. Supported capabilities include function calling, streaming, structured outputs, batch processing, web search, hosted shell, and skills in supported API workflows.
Function calling allows an application to expose defined operations to the model. The model can request one of those operations with structured arguments, while the application remains responsible for executing it and returning the result. This is useful for repository tools, test runners, issue trackers, deployment systems, or internal business operations.
Web search can provide newer information during execution, and hosted shell or other supported tools can help with technical tasks. These tools do not change the model's underlying knowledge cutoff, which OpenAI specifies as August 31, 2025. External information retrieved during a session should therefore be distinguished from information already encoded in the model.
Streaming is available when an application wants partial response delivery rather than waiting for the complete response. Batch processing is also documented, which can be useful for workloads that do not require an immediate interactive result.
Input and output modalities
GPT-5.3-Codex accepts text and image input and produces text output. Image input can be useful for visual debugging, interface analysis, screenshots, diagrams, and computer-use workflows when the surrounding agent supplies the relevant images.
The model does not natively generate images, audio, or video. A request for an image asset, spoken audio, or a video should use a separate specialized capability rather than assuming that GPT-5.3-Codex can produce that media directly. It can still help write code or instructions for a media workflow, but its own model output is text.
GPT-5.3-Codex pricing
OpenAI's documented standard API pricing is:
- Input: $1.75 per 1 million tokens
- Cached input: $0.175 per 1 million tokens
- Output: $14.00 per 1 million tokens
Output tokens cost substantially more than input tokens, so long reasoning traces, extensive generated code, and verbose tool interactions can materially affect total cost. Cached input is priced lower when eligible previously processed content can be reused. The supplied documentation states that fine-tuning is not supported, and predicted outputs are also not supported.
These prices apply to API usage described in the supplied research. The final cost of an application can also depend on how much repository content is sent, how many tool calls are made, whether context is cached, and how often the agent repeats or revises work.
Strengths and trade-offs
The main strength of GPT-5.3-Codex is its fit for complex engineering tasks that combine coding, reasoning, context, and tool use. Its large context window is useful for projects with substantial documentation or many related files. Its reasoning controls allow an application to choose between quicker interaction and more deliberate analysis. Image input adds value when development work includes screenshots or visual interfaces.
The same design creates trade-offs. GPT-5.3-Codex is more expensive than a model chosen purely for inexpensive, high-volume text processing, particularly when output is long. Higher reasoning settings can increase latency. It is also unnecessary for a simple transformation, a short code completion, or a routine classification task where a smaller or faster option would meet the quality requirement.
The model's coding specialization does not eliminate the need for review. Generated patches should be inspected, tests should be run, and commands with production or security consequences should require appropriate authorization. Tool access can extend what the overall agent is able to do, but it also makes permission boundaries and human oversight important.
When to choose GPT-5.3-Codex
Choose GPT-5.3-Codex when the task benefits from several of the following characteristics:
- A large codebase, documentation set, or long-running conversation must remain available as context.
- The work requires repository exploration, multi-file edits, debugging, testing, or iterative refinement.
- The application needs function calling, web search, hosted shell, skills, or other tool-driven workflows.
- More deliberate reasoning is worth additional latency and output cost.
- The model needs to interpret screenshots, diagrams, or other image input while producing text-based technical work.
Consider another model or approach when the task is primarily simple text processing, high-volume low-cost generation, very low-latency interaction, image generation, audio or video generation, or a fine-tuning-dependent application. GPT-5.3-Codex is also not the right choice when the surrounding system cannot safely provide or supervise the tools required for the intended workflow.
Bottom line
GPT-5.3-Codex is best understood as an agentic coding model rather than an ordinary code-completion service. Its 400,000-token context, 128,000-token maximum output, adjustable reasoning effort, image input, and documented tool support make it suitable for extended software engineering tasks. Its higher output price, possible reasoning latency, text-only output, lack of fine-tuning, and need for careful tool supervision are important limitations. For developers building a capable coding agent or handling difficult repository-level work, those trade-offs may be justified; for simpler or cost-sensitive workloads, a smaller or more specialized option may be more appropriate.