Version of GPT-5 optimized for agentic coding in Codex.
GPT-5-Codex is a specialized variant of GPT-5, further optimized for agentic software engineering inside OpenAI's Codex environment. Announced September 23, 2025, it is trained on complex real-world engineering tasks: building full projects from scratch, adding features and tests, debugging, performing large-scale refactors across hundreds of files, and conducting thorough code reviews. On TheRouter, openai/gpt-5-codex is priced identically to GPT-5 at $1.25/$10 per 1M tokens, with 400K context and supports the full set of GPT-5 parameters including reasoning_effort, verbosity, and custom_tools.
Unlike general-purpose GPT-5, GPT-5-Codex dynamically adapts its reasoning time based on task complexity — it produces 93.7% fewer tokens than GPT-5 for simple turns and spends twice as long reasoning on hard problems. It has been observed running independently for over 7 hours on complex tasks. OpenAI recommends GPT-5-Codex only for agentic coding in Codex or Codex-like environments, not as a general-purpose replacement for GPT-5.
| Type | Rate |
|---|---|
| Input | $1.35 / 1M tokens |
| Output | $10.80 / 1M tokens |
| Release date | 2025-09-23openai.com ↗ | verified |
| Training cutoff | Not publicly disclosed | unknown |
| OpenAI model id | gpt-5-codexplatform.openai.com ↗ | verified |
| API availability | Responses API only (natively)openai.com ↗ | verified |
| Dynamic reasoning | 93.7% fewer tokens on simple turns vs GPT-5; 2x reasoning on hard tasksopenai.com ↗ | verified |
| Autonomous execution | Observed >7 hours on complex tasksopenai.com ↗ | verified |
| License | OpenAI Terms of Use (proprietary, API-only) | verified |
| Benchmark | Distribution | Score | Source |
|---|---|---|---|
SWE-bench Verified All 500 tasks. Can reach 82.0% with high-compute config using parallel attempts and internal scoring model. | 77.2%% | openai.com ↗ | |
Code Refactoring Evaluated on refactor tasks from large repos (Python, Go, OCaml). Example: Gitea PR changing 232 files and 3,541 lines. | OpenAI internal eval | openai.com ↗ | |
Code Review Accuracy Fewer incorrect/unimportant comments than GPT-5 per experienced SWE evaluations on popular open-source repos. | Highly preferred vs GPT-5 | 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-codex",
"messages": [
{"role": "user", "content": "Summarize the key points from this input."}
]
}'Via TheRouter, gpt-5-codex is called through the standard chat completions endpoint. TheRouter abstracts the native Responses API to the familiar chat completions interface.
curl https://api.therouter.ai/v1/chat/completions \
-H "Authorization: Bearer $THEROUTER_API_KEY" \
-H "Content-Type: application/json" \
-d '{
"model": "openai/gpt-5-codex",
"messages": [
{"role": "user", "content": "Write a Python script to parse logs and extract error counts for each module."}
],
"reasoning_effort": "medium"
}'GPT-5.5 matches GPT-5.4 per-token latency while delivering significantly higher intelligence and using fewer tokens per Codex task. Ships with OpenAI's strongest safeguards.
GPT-5.3-Codex sets new highs on SWE-Bench Pro and Terminal-Bench. Documents safety evaluations including destructive actions mitigation.
GPT-5.2-Codex achieves SOTA on SWE-Bench Pro and Terminal-Bench 2.0. Optimized for long-horizon agentic coding and large-scale refactors.
GPT-5-Codex announced as a specialized version of GPT-5 for Codex. Available via Responses API at same pricing as GPT-5 ($1.25/$10 per 1M tokens).
GPT-5-Codex is a specialized version of GPT-5, purpose-built for agentic coding in OpenAI's Codex environment. It dynamically adjusts reasoning time based on task complexity — fewer tokens on simple requests, more reasoning on hard ones. OpenAI recommends it only for coding tasks in Codex or Codex-like environments.
Yes. Set model to 'openai/gpt-5-codex' on api.therouter.ai/v1/chat/completions with standard OpenAI SDK. All GPT-5 parameters work: reasoning_effort, verbosity, tools, streaming, structured output.
Identical: $1.25/$10 per 1M tokens, cache_read $0.125. Effective cost per task can be lower because GPT-5-Codex uses fewer tokens on simple requests (93.7% fewer for simplest 10% of turns).
No. OpenAI explicitly recommends GPT-5-Codex only for agentic coding in Codex-like environments. For general reasoning, use openai/gpt-5.
| source | URL | retrieved | |
|---|---|---|---|
| Release date | openai.com ↗ | 2026-05-28 | verified |
| Training cutoff | — | — | unknown |
| OpenAI model id | platform.openai.com ↗ | 2026-05-28 | verified |
| API availability | openai.com ↗ | 2026-05-28 | verified |
| Dynamic reasoning | openai.com ↗ | 2026-05-28 | verified |
| Autonomous execution | openai.com ↗ | 2026-05-28 | verified |
| License | — | — | verified |
| SWE-bench Verified | openai.com ↗ | 2026-05-28 | verified |
| Code Refactoring | openai.com ↗ | 2026-05-28 | verified |
| Code Review Accuracy | openai.com ↗ | 2026-05-28 | verified |
| OpenAI releases GPT-5.5 with stronger Codex performance | openai.com ↗ | 2026-05-28 | verified |
| OpenAI releases GPT-5.3-Codex system card | openai.com ↗ | 2026-05-28 | verified |
| OpenAI launches GPT-5.2-Codex with SOTA on agentic coding benchmarks | openai.com ↗ | 2026-05-28 | verified |
| OpenAI releases GPT-5-Codex for agentic coding in Codex | openai.com ↗ | 2026-05-28 | verified |
| What is the difference between GPT-5 and GPT-5-Codex? | openai.com ↗ | 2026-05-28 | to verify |
| Can I use GPT-5-Codex via TheRouter's API? | openai.com ↗ | 2026-05-28 | to verify |
| How does GPT-5-Codex pricing compare to GPT-5? | openai.com ↗ | 2026-05-28 | to verify |
| Is GPT-5-Codex good for non-coding tasks? | openai.com ↗ | 2026-05-28 | to verify |