GPT-5 Codex

The First GPT-5 Model Purpose-Built for Codex

OpenAI introduced GPT-5 Codex on September 15, 2025 and expanded API-key access on September 23, 2025.

The model was designed for work where coding is an iterative process rather than a single prompt-to-code response.

OpenAI trained GPT-5 Codex on complex real-world engineering tasks including:

• building complete projects from scratch;
• adding features;
• writing tests;
• debugging;
• large-scale refactoring;
• code review;
• repeated test-and-fix loops.

Its role was deliberately narrower than ordinary GPT-5.

OpenAI recommended GPT-5 Codex specifically for agentic coding in Codex or Codex-like environments, while general GPT-5 remained the broader model for non-coding tasks.

The model exposed:

• a 400,000-token context window;
• up to 128,000 output tokens;
• text input and output;
• image input;
• reasoning tokens;
• streaming;
• function calling;
• structured outputs;
• Responses API access.

GPT-5 Codex is no longer available through the API. OpenAI shut it down on July 23, 2026.

The Model Did Not Treat Every Turn the Same

OpenAI reported a large difference in model-generated token usage between easy and difficult work.

For the bottom 10% of internal user turns by generated tokens, GPT-5 Codex used 93.7% fewer tokens than GPT-5.

For the top 10%, GPT-5 Codex spent roughly twice as long reasoning, editing, testing, and iterating.

OpenAI also reported seeing the model work independently for more than seven hours on large, complex tasks during testing.

That behavior is central to the GPT-5 Codex identity.

The model was not simply configured to “reason more.”

It was trained to scale the entire engineering trajectory:

• inspect;
• plan;
• edit;
• test;
• diagnose;
• repair;
• repeat.

For migration, preserve both ends of the workload.

A replacement should be evaluated on:

• quick interactive edits;
• medium implementation tasks;
• long-running refactors;
• tasks with multiple test failures;
• tasks requiring recovery after a wrong approach.

Agent Cost Is Larger Than One Request

Historical API rates were:

• input: $1.25 per 1M tokens;
• cached input: $0.125 per 1M tokens;
• output: $10.00 per 1M tokens.

Reasoning tokens are part of the generated-token economics.

For an agentic coding task, a single headline rate is not enough to estimate cost.

A coding trajectory can include:

• repository context;
• repeated tool results;
• hidden reasoning;
• visible progress;
• generated patches;
• test output;
• repair attempts;
• multiple responses.

That is why historical migration should compare cost per accepted patch or cost per completed task.

A replacement that costs less per token may still cost more if it needs more failed attempts.

A replacement that costs more per token may be cheaper if it completes the task in fewer iterations.

Reasoning Was Part of the Agent Loop

Reasoning tokens let a model plan, inspect alternatives, decide when to use tools, diagnose failures, and determine what to do next.

For coding agents, that reasoning is intertwined with execution.

The model may:

1. inspect the repository;
2. form a hypothesis;
3. edit code;
4. run tests;
5. observe a failure;
6. revise its understanding;
7. edit again;
8. verify completion.

OpenAI’s launch data showed that GPT-5 Codex could dramatically reduce generated tokens on simpler turns while expanding work on difficult ones.

This is a more useful historical distinction than merely labeling the model “high reasoning.”

The agent was optimized for adaptive software-engineering effort.

When evaluating a replacement, preserve the original task difficulty distribution.

Do not benchmark only the hardest repository tasks or only easy edits.

Measure:

• reasoning tokens;
• total generated tokens;
• tool calls;
• elapsed task time;
• repair iterations;
• accepted-result quality.

The API Choice Reflected the Model’s Purpose

The official model page lists GPT-5 Codex as available in the Responses API only.

The model supported:

• streaming;
• function calling;
• structured outputs;
• reasoning tokens;
• text input and output;
• image input.

It did not support:

• audio input or output;
• video input;
• fine-tuning;
• predicted outputs.

The Responses API is better suited to an agent loop because the response can include structured items rather than only a chat message.

Applications can maintain:

• model responses;
• tool calls;
• tool outputs;
• reasoning-related state;
• structured results.

For migration, preserve the Responses API architecture where possible.

OpenAI’s current guidance continues to favor Responses for reasoning models and tool-based agentic workflows.

Preserve the Workflow, Then Change the Model

The original GPT-5 Codex integration was optimized around:

• repository navigation;
• AGENTS.md instructions;
• iterative tool use;
• long-running reasoning;
• code review;
• tests and validation;
• Responses API state.

Keep those parts stable during the first migration evaluation.

Then compare replacement behavior.

Use representative tasks from production:

• small bug fix;
• feature implementation;
• multi-file refactor;
• test repair;
• front-end change using screenshots;
• code review;
• long-running task with repeated failures.

Measure:

• accepted patch rate;
• tests passed;
• reasoning and output tokens;
• commands and tool calls;
• latency;
• number of repair cycles;
• unnecessary edits;
• human corrections;
• total task cost.

Newer model generations also expose different reasoning controls.

Do not copy an old reasoning setting blindly.

Run an effort sweep using the currently supported values for the replacement model and find the cheapest setting that preserves task completion.