OpenAI Codex model
GPT-5.3 Codex
A specialized GPT-5.3 model built for long-running agentic coding, repository-scale engineering, research, tool use, and complex software execution.
- Context window
- 400K
- tokens
- Max output
- 128K
- tokens
- Input
- $1.75
- per 1M tokens
- Cached input
- $0.175
- per 1M tokens
- Output
- $14.00
- per 1M tokens
01 / Overview
What GPT-5.3 Codex Is
GPT-5.3 Codex is a specialized GPT-5.3 model optimized for agentic coding in Codex and similar environments, combining software-engineering capability with professional reasoning, research, tool use, and long-running execution.
Built to Keep Working After the First Code Change
Traditional code generation focuses on producing a function, snippet, or patch from one prompt. Agentic coding is broader. The model may need to inspect a repository, understand existing conventions, search for relevant files, form a plan, edit several locations, run commands, inspect failures, revise its approach, and continue until the task is complete.
GPT-5.3 Codex was built around that workflow. OpenAI introduced it as its most capable agentic coding model at the time, combining improvements from GPT-5.2 Codex with the reasoning and professional knowledge capabilities of GPT-5.2.
OpenAI also reported that GPT-5.3 Codex was approximately 25% faster than its predecessor and could handle long-running work involving research, tool use, and complex execution. In Codex environments, users could steer and interact with the model while it was working without discarding task context.
The model is now deprecated. OpenAI announced deprecation on October 1, 2026, with API shutdown scheduled for April 1, 2027. GPT-6 Sol is the recommended replacement.
- Specialized for agentic coding rather than generic chat.
- Designed for long-running software-engineering tasks.
- Combines coding, research, tools, and professional reasoning.
- Supports a 400K-token context window.
- Deprecated with GPT-6 Sol as the recommended replacement.
- Provider
- OpenAI
- Family
- GPT-5.3
- Specialization
- Agentic coding
- Status
- Deprecated
- Knowledge cutoff
- Aug 31, 2025
- Input modalities
- Text, Image
- Reasoning
- Low to XHigh
02 / Use cases
Where GPT-5.3 Codex Fits Best
GPT-5.3 Codex was designed for engineering tasks where success depends on navigating a codebase and executing a sequence of actions, not merely generating code in isolation.
Repository Work, Research, and Complex Execution
Large repository tasks are a natural fit. A coding agent may need to locate the implementation behind a feature, understand how several modules interact, identify tests and conventions, and apply coordinated changes without breaking unrelated behavior.
Debugging is another agentic workload. The model can search source files, inspect command output, formulate a hypothesis, modify code, and continue iterating based on the result.
OpenAI also positioned GPT-5.3 Codex beyond pure code generation. Its launch emphasized research, tool use, and complex execution across professional work on a computer. That makes the model historically relevant to engineering workflows that mix code with documentation, investigation, and operational actions.
For migration, these complete workflows are more useful than isolated coding prompts. A replacement should be evaluated on whether it finishes the same tasks with comparable correctness, tool discipline, autonomy, and token cost.
- Repository-scale feature implementation.
- Multi-file debugging and iterative repair.
- Codebase exploration and technical investigation.
- Tool-driven engineering tasks with shell execution.
- Long-running work that combines research and code changes.
- 01
Repository navigation
Search files, understand project structure, trace behavior, and identify the code that actually needs to change.
- 02
Long-running implementation
Carry a software task through planning, editing, command execution, verification, and follow-up fixes.
- 03
Debugging loops
Inspect failures, develop hypotheses, change code, rerun tools, and iterate toward a working result.
- 04
Research-assisted engineering
Combine codebase context, external information, and tools when implementation depends on more than local source code.
03 / Pricing
GPT-5.3 Codex Pricing
GPT-5.3 Codex is listed at $1.75 per million input tokens, $0.175 per million cached input tokens, and $14.00 per million output tokens.
Agent Cost Is Shaped by the Entire Engineering Trajectory
Coding agents often make several model turns to complete one user-visible task. Repository context, tool output, test logs, intermediate decisions, and generated patches can all contribute to total usage.
That means the useful metric is usually cost per completed engineering task rather than cost per API request.
Cached input can reduce the price of repeated stable prefixes such as repository instructions, system prompts, and other reusable context. Output remains substantially more expensive than input, so verbose plans, explanations, and repeated patch generation can materially affect total spend.
The preserved token rates are also useful as a migration baseline. Teams can compare the full cost of historical GPT-5.3 Codex tasks against GPT-6 Sol instead of comparing headline prices without accounting for task completion and retry rates.
- $1.75 per 1M input tokens.
- $0.175 per 1M cached input tokens.
- $14.00 per 1M output tokens.
- Tool-specific usage may introduce separate fees.
- Evaluate migration using cost per completed engineering task.
1M tokens · USD
- Input
- $1.75
- Cached input
- $0.175
- Output
- $14.00
Example: 10K input + 2K output
- Input cost
- $0.0175
- Output cost
- $0.0280
- Estimated total
- $0.0455
04 / Context
A 400K Context Window for Repository-Scale Work
GPT-5.3 Codex provides a 400,000-token context window and supports up to 128,000 output tokens, giving coding agents substantial space for source files, project instructions, task history, and tool results.
Context Is the Agent's Working Memory
A coding agent needs more than the file it is editing. It may require repository-level instructions, neighboring implementations, interfaces, tests, configuration, issue context, command output, and earlier decisions made during the same task.
The 400K context window gives GPT-5.3 Codex room to maintain a substantial working set without requiring every engineering task to be fragmented into tiny independent requests.
Still, more context is not automatically better. Large repositories can contain many files that do not matter to the current task. Effective agent systems search and select relevant information rather than placing an entire codebase into every request.
Migration testing should preserve the same context strategy initially. If both the model and retrieval strategy change at once, it becomes harder to identify whether differences in coding quality came from the model or from the information it received.
- 400,000-token context window.
- Maximum output of 128,000 tokens.
- Suitable for substantial repository working sets.
- Can retain task history and tool results across complex work.
- Retrieval and file selection still matter for efficiency and accuracy.
Context window
400,000
Max output
128,000
A coding-agent context can include repository instructions, selected source files, tests, issue requirements, command output, research results, and intermediate task state.
05 / Reasoning
Four Reasoning Levels for Coding Tasks
GPT-5.3 Codex supports low, medium, high, and xhigh reasoning effort, allowing coding systems to trade inference depth against latency and token usage.
Match Reasoning Depth to Engineering Complexity
A straightforward repository search or targeted edit may not require the same reasoning budget as a cross-cutting architectural change or difficult debugging investigation.
The low and medium settings give developers lower-cost baselines for bounded coding work. high and xhigh are better candidates for tasks where the model must reason across several dependencies, diagnose uncertain failures, or coordinate a longer sequence of actions.
OpenAI's GPT-5.3 Codex guidance also emphasizes autonomy and follow-through. For agent evaluations, reasoning effort should be tested together with tool instructions and execution policy rather than in isolation.
There is no need to assume one setting is universally best. A useful benchmark separates task classes and finds the lowest reasoning effort that reliably completes each class.
Bounded coding task
Test low or medium for repository search, focused edits, and well-specified implementation work.
Long-horizon engineering
Test high or xhigh for difficult debugging, architectural changes, research-heavy tasks, and multi-step execution.
06 / Capabilities
Capabilities for Agentic Coding Environments
GPT-5.3 Codex supports text and image input, streaming, function calling, structured outputs, and a Codex-oriented tool workflow that OpenAI documents around web search, hosted shell, Skills, and application-defined functions.
Designed to Reason and Act Inside an Engineering Loop
The model accepts text and image input and produces text output. Image input can be useful when a coding task involves screenshots, visual bugs, interface references, or design artifacts.
Streaming is supported, allowing an application to expose progress or model output incrementally.
Function calling and structured outputs provide application-level control for workflows that need predictable machine-readable results or custom actions.
OpenAI's dedicated GPT-5.3 Codex API guidance specifically highlights function calling, web search, hosted shell, and Skills. Hosted shell is especially relevant to coding agents because it allows a workflow to inspect files, run commands, execute tests, and use the result as part of the next reasoning step.
Fine-tuning and predicted outputs are not supported.
- Text input and output with image input.
- Streaming supported.
- Function calling supported.
- Structured outputs supported.
- Web search supported in documented Codex workflows.
- Hosted shell supported in documented Codex workflows.
- Skills supported in documented Codex workflows.
- Fine-tuning is not supported.
- Supported
Hosted shell
Run command-line operations and feed execution results back into a coding workflow.
- Supported
Web search
Retrieve external information when implementation requires research beyond the repository.
- Supported
Skills
Use reusable instructions and workflows in Codex-style agent environments.
- Supported
Function calling
Connect model decisions to application-defined engineering actions.
- Supported
Structured outputs
Return machine-readable results when the surrounding system requires a schema.
- Supported
Image input
Include screenshots and visual references alongside text and code context.
07 / Evaluation
Strengths and Limitations
GPT-5.3 Codex was a major agentic-coding model for long-running software work, but its deprecated status now makes it primarily a migration baseline for teams moving coding agents to GPT-6 Sol.
Historical strengths
Agentic coding specialization
The model was optimized specifically for coding tasks that require repository navigation, tools, iteration, and sustained execution.
Long-running task capability
OpenAI designed GPT-5.3 Codex for work that combines research, tool use, and complex multi-step execution.
Large coding context
A 400K-token context gives agents room for substantial source files, project instructions, logs, and task state.
Configurable reasoning
Low through xhigh reasoning levels allow different engineering tasks to use different amounts of inference effort.
What to consider
Deprecated lifecycle
GPT-5.3 Codex was deprecated on October 1, 2026 and is scheduled to shut down on April 1, 2027. OpenAI recommends GPT-6 Sol.
Specialized rather than general-purpose
The model was tuned for Codex and similar agentic coding environments rather than broad conversational applications.
Newer coding models are available
Current deployments should evaluate GPT-6 Sol and other supported models instead of starting new production integrations on GPT-5.3 Codex.
Fine-tuning unavailable
OpenAI lists fine-tuning as unsupported for GPT-5.3 Codex.
Preserve coding-agent behavior
Benchmark GPT-5.3 Codex before migrating your coding workflows
Capture representative repository tasks, tool calls, patching workflows, and long-running engineering cases, then compare quality, token usage, and cost against a supported replacement in EidoStack.
Start FreeGPT-5.3 Codex is deprecated. OpenAI recommends migrating to GPT-6 Sol before the April 1, 2027 API shutdown.
Common Questions
What is GPT-5.3 Codex?
GPT-5.3 Codex is an OpenAI model optimized for agentic coding in Codex and similar environments. It was designed for long-running engineering tasks involving repository navigation, research, tools, and complex execution.
Is GPT-5.3 Codex deprecated?
Yes. OpenAI deprecated GPT-5.3 Codex on October 1, 2026 and plans to remove it from the API on April 1, 2027.
What should replace GPT-5.3 Codex?
OpenAI lists GPT-6 Sol as the recommended replacement for GPT-5.3 Codex.
How much does GPT-5.3 Codex cost?
The model card lists $1.75 per 1M input tokens, $0.175 per 1M cached input tokens, and $14.00 per 1M output tokens.
What is the context window of GPT-5.3 Codex?
GPT-5.3 Codex has a 400,000-token context window and supports up to 128,000 output tokens.
What reasoning levels does GPT-5.3 Codex support?
GPT-5.3 Codex supports low, medium, high, and xhigh reasoning effort.
What is the knowledge cutoff for GPT-5.3 Codex?
OpenAI lists August 31, 2025 as the knowledge cutoff for GPT-5.3 Codex.
Does GPT-5.3 Codex support image input?
Yes. GPT-5.3 Codex accepts text and image input and generates text output. Audio and video are not supported.
Does GPT-5.3 Codex support streaming?
Yes. OpenAI lists streaming as supported.
Does GPT-5.3 Codex support function calling and structured outputs?
Yes. OpenAI lists both function calling and structured outputs as supported.
What tools are documented for GPT-5.3 Codex?
OpenAI's dedicated GPT-5.3 Codex guide highlights function calling, web search, hosted shell, and Skills for agentic coding workflows.
What workloads are a good fit for GPT-5.3 Codex?
Historically, the model was a strong fit for repository-scale implementation, multi-file debugging, codebase exploration, long-running software tasks, and research-assisted engineering. New deployments should use a supported replacement.
Can GPT-5.3 Codex be fine-tuned?
No. OpenAI lists fine-tuning as unsupported for GPT-5.3 Codex.
Model information
Last updated
The specifications, pricing, capabilities, and lifecycle information on this page are based on the official OpenAI GPT-5.3 Codex model card, release announcement, API guide, changelog, and deprecation documentation. Provider behavior and migration guidance can change, so verify current production assumptions against OpenAI's latest documentation.