GPT-5.3-Codex
GPT-5.3-Codex by OpenAI is a Codex-native agent merging frontier coding with GPT-5.2 reasoning, real-time steering, computer use, and 400K-token context.
Model Overview
Capabilities, design details, and architectural traits
GPT-5.3-Codex - The First Model Instrumental in Its Own Creation
GPT-5.3-Codex is OpenAI's most capable agentic coding model, and the first in the GPT 5 family to directly combine frontier coding performance with professional knowledge and general reasoning in a single model. Uniquely, early versions were used by the Codex team to debug their own training run, manage deployment, and diagnose evaluations - making GPT-5.3-Codex the first model documented as having been instrumental in creating itself.
What Makes GPT-5.3-Codex Different
| Trait | Detail |
|---|---|
| Merged training stacks | First model to combine the GPT-5.2-Codex coding stack with the GPT-5.2 reasoning and professional knowledge stack, rather than specializing in one |
| Real-time steering | Users can redirect the agent mid-task without losing context - described as interacting "much like a colleague" while it works |
| Full software lifecycle scope | Targets debugging, deploying, monitoring, writing PRDs, editing copy, user research, metrics, slide decks, and spreadsheets - not code generation only |
| High-capability cybersecurity classification | First OpenAI model classified as High capability for cybersecurity under the Preparedness Framework; trained to identify software vulnerabilities |
| Reasoning effort control | Supports low, medium, high, and xhigh effort settings, allowing per-request compute tuning in Codex and API environments |
| Self-referential development loop | Used to optimize its own training harness, root-cause inference bugs, and analyze its own session logs |
Codex-Native Context
GPT-5.3-Codex is designed for the Codex environment - the app, CLI, and IDE extension - and available via both Chat Completions and Responses API. Its 400k token context window and 128k token max output support long-horizon tasks that iterate over millions of tokens, such as autonomously building multi-map games or full financial presentation decks from a single prompt.
Benchmark Performance
Independent evaluations · Artificial Analysis
Accuracy & Capability Details
Compare Models Side-by-Side
Evaluate specifications, pricing, and independent benchmark indices
| Model Details | |||
|---|---|---|---|
| General Info | |||
| Provider | OpenAI | Anthropic | Anthropic |
| Release Date | February 5, 2026 | September 22, 2026 | September 28, 2026 |
| Knowledge Cutoff | - | - | Jun 2026 |
| Context & Limits | |||
| Context Window | 400K | 1M Best Context Window | 1M Best Context Window |
| Pricing (per 1M tokens) | |||
| Input Pricing | $1.75 Best Input Pricing | $4 | $2 |
| Output Pricing | $14 | $20 | $10 Best Output Pricing |
| Modalities | |||
| Inputs | textimagefile | textimagefile | textimagefile |
| Outputs | text | text | text |
| Benchmarks (0-100) | |||
| Intelligence Index | 32.5 | 57.6 Best Intelligence Index | 56.0 |
| Coding Index | - | - | - |
| Agentic Index | - | - | - |
Humanity's Last Exam
Extremely difficult logical reasoning and knowledge.
Long Context Reasoning
Logical reasoning over long context windows.
SciCode Benchmark
Scientific coding and mathematical modeling.
Independent evaluation data provided by Artificial Analysis. To view the latest benchmarks and full details, visit their official site.
Explore more from OpenAI
Other models by OpenAI
Top AI Models
Leading alternatives by intelligence score
