OpenAI's intelligent reasoning model for coding and agentic tasks with configurable reasoning effort. Features a 400K context window and 128K max output.
GPT-5 is OpenAI's first unified reasoning model, launched August 7, 2025, succeeding GPT-4o and o3. It is the founding model of the GPT-5 family β subsequent iterations (GPT-5.1 through GPT-5.5) have since shipped. Described by OpenAI as a single system that knows when to answer quickly and when to think longer, GPT-5 combines a fast base model, a deep reasoning model (GPT-5 thinking), and a real-time router that selects between them based on conversation complexity. On TheRouter, the API-accessible 'gpt-5' is the reasoning model that powers maximum performance in ChatGPT, available with 400K context, 128K max output, and $1.25 / $10 per 1M tokens.
From a TheRouter operator's perspective, the original gpt-5 is stable but has been superseded by gpt-5.2, gpt-5.4, and gpt-5.5 which incorporate iterative improvements. The model ships with a new verbosity parameter (low, medium, high), reasoning_effort can be set to minimal for faster responses, and a new custom_tools type supports plaintext tool definitions constrained by developer-supplied context-free grammars. GPT-5 was widely adopted by coding agents (Cursor, Windsurf, GitHub Copilot) and set state-of-the-art on SWE-bench Verified (74.9%), Aider Polyglot (88%), and Ο2-bench telecom (96.7%) at launch.
| Type | Rate |
|---|---|
| Input | $1.35 / 1M tokens |
| Output | $10.80 / 1M tokens |
| Release date | 2025-08-07openai.com β | verified |
| OpenAI model id | gpt-5openai.com/dev β | verified |
| Training cutoff | Not publicly disclosed | unknown |
| Reasoning effort levels | minimal, low, medium, high (API default)openai.com/dev β | verified |
| Verbosity levels | low, medium, high (API default medium)openai.com/dev β | verified |
| Custom tool type support | Custom tools via plaintext + context-free grammarsopenai.com/dev β | verified |
| Status in GPT-5 family | Original release, superseded by GPT-5.1+help.openai.com β | verified |
| License | OpenAI Terms of Use (proprietary, API-only) | verified |
| Benchmark | Distribution | Score | Source |
|---|---|---|---|
SWE-bench Verified Reasoning model at high effort. Uses 22% fewer output tokens and 45% fewer tool calls than o3 at equivalent setting. | 74.9%% | openai.com/dev β | |
Aider Polyglot Code editing benchmark (diff format). One-third reduction in error rate compared to o3. | 88%% | openai.com/dev β | |
AIME 2025 Without tools. State-of-the-art on the American Invitational Mathematics Examination. | 94.6%% | openai.com β | |
MMMU Massive Multi-discipline Multimodal Understanding β college-level vision + language reasoning. | 84.2%% | openai.com β | |
Ο2-bench (telecom) Tool-calling benchmark for agentic tasks. SOTA at launch on this benchmark released June 2025. | 96.7%% | openai.com/dev β | |
HealthBench Hard Health reasoning benchmark based on physician-defined criteria. Highest score of any model at launch. | 46.2%% | openai.com β | |
GPQA Graduate-level QA. Set SOTA without tools (GPT-5 pro extended reasoning). | 88.4%% | openai.com β | |
Scale MultiChallenge Multi-turn instruction following, graded by o3-mini. Tests four types of information use across conversation turns. | 69.6%% | openai.com/dev β |
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",
"messages": [
{"role": "user", "content": "Summarize the key points from this input."}
]
}'Standard chat completion. GPT-5 supports the full OpenAI chat interface with reasoning_effort and verbosity parameters.
curl https://api.therouter.ai/v1/chat/completions \
-H "Authorization: Bearer $THEROUTER_API_KEY" \
-H "Content-Type: application/json" \
-d '{
"model": "openai/gpt-5",
"messages": [{"role": "user", "content": "Explain how multi-agent coordination works"}],
"reasoning_effort": "medium",
"verbosity": "high"
}'GPT-5 launched as OpenAI's first unified reasoning system combining a fast base model, deep reasoning model, and real-time router. SOTA across AIME 2025 (94.6%), SWE-bench Verified (74.9%), and MMMU (84.2%).
OpenAI retired GPT-5 Instant and GPT-5 Thinking from ChatGPT in a spring cleanup. The API model gpt-5 remains available.
The API model gpt-5 is the max-performance reasoning model powering ChatGPT's Thinking mode. ChatGPT uses a unified system with a fast base model (gpt-5-chat-latest), reasoning model (gpt-5), and internal router. The API exposes gpt-5 directly with configurable reasoning_effort.
Each iteration adds improvements in coding, agentic capabilities, and response quality. The original gpt-5 (Aug 2025) is the foundation; later versions incorporate community feedback and production learnings. For new projects, consider the latest stable iteration. The original gpt-5 remains available for backward compatibility.
custom_tools is a GPT-5 feature for defining tools via plaintext descriptions instead of JSON Schema, optionally constrained by context-free grammars. Standard function calling (type: function) still works and is recommended for most cases.
| source | URL | retrieved | |
|---|---|---|---|
| Release date | openai.com β | 2026-05-28 | verified |
| OpenAI model id | openai.com/dev β | 2026-05-28 | verified |
| Training cutoff | β | β | unknown |
| Reasoning effort levels | openai.com/dev β | 2026-05-28 | verified |
| Verbosity levels | openai.com/dev β | 2026-05-28 | verified |
| Custom tool type support | openai.com/dev β | 2026-05-28 | verified |
| Status in GPT-5 family | help.openai.com β | 2026-05-28 | verified |
| License | β | β | verified |
| SWE-bench Verified | openai.com/dev β | 2026-05-28 | verified |
| Aider Polyglot | openai.com/dev β | 2026-05-28 | verified |
| AIME 2025 | openai.com β | 2026-05-28 | verified |
| MMMU | openai.com β | 2026-05-28 | verified |
| Ο2-bench (telecom) | openai.com/dev β | 2026-05-28 | verified |
| HealthBench Hard | openai.com β | 2026-05-28 | verified |
| GPQA | openai.com β | 2026-05-28 | verified |
| Scale MultiChallenge | openai.com/dev β | 2026-05-28 | verified |
| OpenAI launches GPT-5 | openai.com β | 2026-05-28 | verified |
| GPT-5 (Instant/Thinking) retired in ChatGPT | help.openai.com β | 2026-05-28 | verified |
| What's the difference between gpt-5 in the API and GPT-5 in ChatGPT? | openai.com β | 2026-05-28 | to verify |
| How does original gpt-5 differ from GPT-5.2 / 5.4 / 5.5? | help.openai.com β | 2026-05-28 | to verify |
| What are custom_tools and how do they work? | openai.com β | 2026-05-28 | to verify |