OpenAI Codex

Total Parameters: 0

Category	Introduced / Model	Parameter	Syntax / Key	Accepted Values / Range	Code Example	Description & Notes
1. Core Model	All Models	Model ID	`model`	gpt-4o, gpt-4o-mini, o3-mini, o1, code-davinci-002	`"model": "gpt-4o"`	Selects the underlying engine. Use `gpt-4o` for production coding or `o3-mini` for complex algorithmic reasoning.
1. Core Model	All Models	Temperature	`temperature`	0.0 – 2.0 (Default 1.0; 0.0 for code)	`"temperature": 0.0`	Controls randomness. Use `0.0` for deterministic, syntax-exact code generation and math tasks.
1. Core Model	All Models	Max Tokens	`max_completion_tokens`	1 – 128,000 (Model dependent)	`"max_completion_tokens": 2048`	Sets the maximum number of output tokens (code + reasoning) generated in a single completion response.
1. Core Model	All Models	Top P	`top_p`	0.0 – 1.0 (Default 1.0)	`"top_p": 0.1`	Nucleus sampling alternative to temperature. `0.1` means only top 10% probability mass tokens are evaluated.
2. Code Control	Modern API	Response Format	`response_format`	json_object, json_schema, text	`"type": "json_schema"`	Guarantees output conforms strictly to a JSON Schema or valid JSON object. Ideal for automated AST/code parsing.
2. Code Control	All Models	Stop Sequences	`stop`	String or Array (Up to 4 strings)	"stop": ["\nclass ", "```"]	Halts code generation immediately when specified tokens are encountered (e.g. stop at function/class boundary).
2. Code Control	Function Calling	Tools / Functions	`tools`	Array of function objects	`"type": "function"`	Enables Codex to output structured call arguments to run sandbox tools, execute Python, or invoke external APIs.
2. Code Control	Reproducibility	Seed	`seed`	Integer (e.g., 42, 1337)	`"seed": 42`	Enables deterministic sampling for repeatable test runs and consistent code benchmarking.
3. Optimization	o-series	Reasoning Effort	`reasoning_effort`	low, medium, high (Default medium)	`"reasoning_effort": "high"`	Controls internal chain-of-thought depth for `o1` / `o3-mini` models before outputting final code.
3. Optimization	All Models	Frequency Penalty	`frequency_penalty`	-2.0 – 2.0 (Default 0.0)	`"frequency_penalty": 0.5`	Penalizes tokens based on their existing frequency in output. Prevents repetitive code loops and infinite loops.
3. Optimization	All Models	Presence Penalty	`presence_penalty`	-2.0 – 2.0 (Default 0.0)	`"presence_penalty": 0.3`	Encourages the model to introduce new programming concepts and functions rather than echoing existing text.
3. Optimization	All Models	Logit Bias	`logit_bias`	Map of Token IDs to values (-100 to 100)	`"logit_bias": {"1234": -100}`	Forces or bans specific tokens from code output (e.g. banning deprecated library calls or unwanted keywords).
4. Codex Infill	Codex Infill	Suffix (FIM)	`suffix`	String (Code coming after cursor)	`"suffix": "\n return result"`	Enables Fill-In-the-Middle (FIM) code completion by providing trailing context after the insertion point.
4. System Roles	All Models	System Message	`role: "system"`	Developer system prompt	`"role": "developer"`	Sets global coding guidelines, language constraints (e.g., Python 3.12, strict TypeScript), and coding style.

Codex Prompt Architecture & Vocabulary

[SYSTEM DIRECTIVE] + [CONTEXT & IMPORTS] + [DOCSTRING / GOAL] + [SIGNATURE / SPEC] => EXECUTABLE CODE

1. Context & Environment

Define language versions, core libraries, and target runtimes (e.g., Python 3.12, AsyncPG, Pydantic v2).

2. Interface Definition

Provide function signatures, class scaffolding, or TypeScript interfaces to anchor structural intent.

3. Guardrails & Limits

Specify error handling strategies, edge cases, performance constraints, and forbidden operations.

4. Expected Output

Request raw executable code without conversational commentary using explicit Markdown code blocks.

Infill & Completion Patterns

FIM (Fill-In-the-Middle) Docstring Completion Few-Shot Code Prompting Zero-Shot Generation

Code Quality Modalities

AST Parsing Unit Test Generation Refactoring & Modernization Security Audit

Codex Power Tips & Best Practices

Zero Temperature for Deterministic Code

When generating production logic, always pass temperature: 0.0. This prevents creative syntax variations and enforces reliable, repeatable code execution. "temperature": 0.0 Tip: Use higher temperature (0.7+) only when generating alternative algorithms.

Docstring-First Prompting

Codex generates significantly higher quality code when given a detailed docstring with type annotations before the code block.

def parse_jwt(token: str) -> dict:
    """Extracts claims safely..."""

Tip: Include @param and @return directives inside comments.

Stop Sequences for Clean Truncation

Prevent Codex from adding extraneous conversational explanations by setting stop sequences on function or class headers. "stop": ["\ndef ", "\nclass ", "```"] Tip: Keeps response tokens minimal and prevents multi-function spillover.

Prompt Framework & Production Example

Core Rules for OpenAI Codex Generation

Specify Exact Versions: Always state library versions (e.g., Pydantic v2, React 18 ESM) to prevent deprecated API syntax.
Provide Input/Output Examples: Include 1–2 JSON or type examples showing expected input payloads and returned outputs.
Enforce Type Safety: Require strict typing (e.g. TypeScript interfaces, Python type hints) for immediate static analysis validation.
Test-Driven Construction (TDD): Provide unit test cases first and ask Codex to write code that satisfies all assertions.

Standard Production Code Prompt Template

// SYSTEM DIRECTIVE: You are a Senior Backend Engineer writing production Python 3.12 code. // REQ: Generate an async function with full typing, error handling, and docstrings. import asyncio from pydantic import BaseModel, EmailStr class UserRegistration(BaseModel): username: str email: EmailStr async def register_user(payload: UserRegistration) -> dict: """ Validates user credentials, creates a database record, and sends a welcome event. @param payload: Validated Pydantic user registration schema. @return: Dict containing user_id and creation status. @raises ValueError: If user already exists in datastore. """ # CODE CONTINUATION HERE...