Most capable agentic coding model to date.
GPT-5.3-Codex is OpenAI's most capable agentic coding model, released February 5, 2026. It merges the frontier coding performance of GPT-5.2-Codex with the reasoning and professional knowledge capabilities of GPT-5.2, offering a 25% speed improvement over its predecessor. OpenAI describes it as the first model that was instrumental in creating itself — the Codex team used early versions to debug training, manage deployment, and diagnose evaluation results.
From a TheRouter operator's perspective, this model is purpose-built for long-horizon agentic coding tasks — multi-file refactors, autonomous bug bashing, and full-stack application development from scratch. It supports vision input (text + image), tools/function calling, structured outputs (response_format), and extended reasoning via the reasoning parameter. On TheRouter, it costs $5 per million input tokens and $40 per million output tokens, with a 400K context window — significantly more expensive than GPT-5.2-Codex but delivering roughly 25% faster throughput and state-of-the-art agentic benchmark scores.
| Type | Rate |
|---|---|
| Input | $1.89 / 1M tokens |
| Output | $15.12 / 1M tokens |
| Release date | 2026-02-05openai.com ↗ | verified |
| Training cutoff | Not publicly disclosed | unknown |
| License | Proprietary — OpenAI API Termsopenai.com ↗ | verified |
| Speed improvement over GPT-5.2-Codex | ~25% fasteropenai.com ↗ | verified |
| Cybersecurity risk tier | High (Precautionary — under Preparedness Framework)openai.com ↗ | verified |
| Benchmark | Distribution | Score | Source |
|---|---|---|---|
SWE-Bench Pro Multi-language rigorous evaluation; exact % not publicly disclosed at launch | State-of-the-art | openai.com ↗ | |
Terminal-Bench 2.0 State-of-the-art; achieved with fewer tokens than any prior model | 77.3%% | llm-stats.com ↗ | |
OSWorld-Verified Computer-use benchmark; human baseline ~72% | 64.7%% | llm-stats.com ↗ | |
SWE-Lancer IC Diamond | 81.4%% | llm-stats.com ↗ | |
GDPval 44 occupations tested; real-world knowledge work tasks | Matches GPT-5.2 | openai.com ↗ |
Use the global api.therouter.ai endpoint shown below for new integrations; the legacy China accelerated endpoint is retired.
curl https://api.therouter.ai/v1/chat/completions -H "Content-Type: application/json" -H "Authorization: Bearer $THE_ROUTER_API_KEY" -d '{
"model": "openai/gpt-5.3-codex",
"messages": [
{"role": "user", "content": "Summarize the key points from this input."}
]
}'Standard chat completions endpoint. GPT-5.3-Codex excels at open-ended agentic coding conversations.
curl https://api.therouter.ai/v1/chat/completions \
-H "Authorization: Bearer $THEROUTER_API_KEY" \
-H "Content-Type: application/json" \
-d '{"model":"openai/gpt-5.3-codex","messages":[{"role":"user","content":"Write a Python script to monitor system CPU and memory usage"}]}'OpenAI announced GPT-5.3-Codex, combining GPT-5.2-Codex coding performance with GPT-5.2 reasoning. Sets SOTA on SWE-Bench Pro and Terminal-Bench 2.0 while being ~25% faster than its predecessor. Notably, it is the first OpenAI model that helped build itself — early versions debugged its own training and deployment.
GPT-5.3-Codex merges the coding capabilities of GPT-5.2-Codex with the reasoning and professional knowledge of GPT-5.2 into a single model, making it more versatile for tasks beyond pure code generation — like research, tool use, and complex execution. It is approximately 25% faster than GPT-5.2-Codex and sets state-of-the-art results on Terminal-Bench 2.0 and SWE-Bench Pro. However, on TheRouter it costs $5/$40 per 1M tokens vs $1.25/$10 for GPT-5.2-Codex, so cost-sensitive routes should compare the two based on task complexity.
Yes. GPT-5.3-Codex accepts both text and image inputs (text + image modality per standard-models.yaml). It does not generate images — output is text only. For image analysis tasks like screenshots of bugs, diagrams, or UI mockups, the model can process those as part of its agentic coding workflow.
| source | URL | retrieved | |
|---|---|---|---|
| Release date | openai.com ↗ | 2026-05-28 | verified |
| Training cutoff | — | — | unknown |
| License | openai.com ↗ | 2026-05-28 | verified |
| Speed improvement over GPT-5.2-Codex | openai.com ↗ | 2026-05-28 | verified |
| Cybersecurity risk tier | openai.com ↗ | 2026-05-28 | verified |
| SWE-Bench Pro | openai.com ↗ | 2026-05-28 | verified |
| Terminal-Bench 2.0 | llm-stats.com ↗ | 2026-05-28 | to verify |
| OSWorld-Verified | llm-stats.com ↗ | 2026-05-28 | to verify |
| SWE-Lancer IC Diamond | llm-stats.com ↗ | 2026-05-28 | to verify |
| GDPval | openai.com ↗ | 2026-05-28 | verified |
| OpenAI introduces GPT-5.3-Codex — the most capable agentic coding model | openai.com ↗ | 2026-05-28 | verified |
| How does GPT-5.3-Codex differ from GPT-5.2-Codex? | openai.com ↗ | 2026-05-28 | to verify |
| Does GPT-5.3-Codex support vision? | openai.com ↗ | 2026-05-28 | to verify |