# GPT-5 ![Organization](https://img.shields.io/badge/Org-OpenAI-412991) ![Type](https://img.shields.io/badge/Type-Reasoning%20system-555) ![Context](https://img.shields.io/badge/Context-400K%20tokens-1f6feb) ![Max output](https://img.shields.io/badge/Max%20output-128K-1f6feb) ![Pricing](https://img.shields.io/badge/Price-%241.25%20%2F%20%2410%20per%201M-2da44e) ![License](https://img.shields.io/badge/License-Proprietary-d29922) ![Modality](https://img.shields.io/badge/Modality-Text%20%2B%20Image-8957e5) ![Released](https://img.shields.io/badge/Released-Aug%202025-555) > **Quick answer:** GPT-5 is OpenAI's flagship large language model, released on 7 August 2025. It is a unified system that pairs a fast everyday model with a deeper "GPT-5 thinking" reasoning model and a real-time router that picks between them. It has a 400,000-token context window and is priced at $1.25 per million input tokens and $10 per million output tokens. On OpenAI's own evaluations it reaches 74.9% on SWE-bench Verified and 94.6% on AIME 2025 (without tools). It is a closed, API-only model. ## At a Glance **Where GPT-5 leads** - Frontier agentic coding and reasoning: 74.9% on SWE-bench Verified and 94.6% on AIME 2025 (no tools). - Aggressive pricing for a frontier model — $1.25 input / $10 output per 1M tokens, well below Claude Opus 4.6 and most rivals. - A router that spends expensive reasoning tokens only when a query needs them, keeping everyday calls fast and cheap. **Where it lags** - Its 400K context window trails the 1M-token windows of Claude Opus 4.6 and Gemini 3 Pro. - Multimodality is limited to text and image input — no audio or video — and there are no open weights to self-host. **Best for:** teams that want frontier-level coding and agent performance at the lowest frontier price. ## What GPT-5 Is GPT-5 is OpenAI's most capable model family and the default in ChatGPT since August 2025. Rather than a single network, it ships as a unified system: a fast, efficient model handles most queries, a deeper reasoning model ("GPT-5 thinking") takes harder problems, and a continuously trained router decides which to use based on the conversation's complexity, tool needs, and explicit cues like "think hard about this." When usage limits are hit, a mini variant of each model serves the remainder. For the most demanding work, OpenAI also offers GPT-5 pro — a variant that spends far longer on parallel test-time compute to produce the highest-quality answers in the family, and which holds OpenAI's reported state of the art on GPQA. The whole system was trained on Microsoft Azure AI supercomputers and is multimodal, reasoning over images, charts, and diagrams in addition to text. Because GPT-5 is closed and available only through OpenAI's API and ChatGPT, Benchgen treats it as an API-benchmarked model: there are no open weights to host, so its value on the platform comes from running it through Benchgen's evaluations via API calls and comparing the verified results against its strong but general public scores. The interesting question is not whether GPT-5 is capable in the abstract — it clearly is — but how reliably that capability converts into completed, domain-specific agent tasks across repeated trials. ## Specifications | Field | Value | |---|---| | Organization | OpenAI | | Model type | Unified reasoning system (fast model + GPT-5 thinking + router; GPT-5 pro variant) | | API identifier | `gpt-5-2025-08-07` | | Access | Closed, API only (also default in ChatGPT) | | Context window | 400,000 tokens | | Max output | 128,000 tokens | | License | Proprietary | | Release date | 7 August 2025 | | Knowledge cutoff | September 2024 | | Modality | Multimodal (text and image input; text output) | | Training infrastructure | Microsoft Azure AI supercomputers | ## Pricing | | Input (per 1M tokens) | Output (per 1M tokens) | |---|---|---| | OpenAI | $1.25 | $10.00 | Cached input is billed at $0.125 per 1M tokens, and the Batch API discounts both input and output by 50%. Rates shown are standard processing for prompts under 270K tokens. Source: [OpenAI API pricing](https://openai.com/api/pricing/) / [GPT-5 model docs](https://developers.openai.com/api/docs/models/gpt-5). ## Context Window GPT-5 has a **400,000-token context window** — roughly **600 pages** of text in a single request — with a maximum output of **128,000 tokens**. That is enough to hold a large codebase, a long contract set, or dozens of documents at once, while the router decides how much reasoning to apply across that context. ## Public Benchmark Scores | Benchmark | Score | Source | Date | |---|---|---|---| | [SWE-bench Verified](/benchmarks/swe-bench-verified) | 74.9% | [OpenAI — Introducing GPT-5](https://openai.com/index/introducing-gpt-5/) | 2025-08 | | [AIME 2025](/benchmarks/aime-2025) (no tools) | 94.6% | [OpenAI — Introducing GPT-5](https://openai.com/index/introducing-gpt-5/) | 2025-08 | | Aider Polyglot | 88% | [OpenAI — Introducing GPT-5](https://openai.com/index/introducing-gpt-5/) | 2025-08 | | MMMU | 84.2% | [OpenAI — Introducing GPT-5](https://openai.com/index/introducing-gpt-5/) | 2025-08 | | [GPQA](/benchmarks/gpqa-diamond) (GPT-5 pro, no tools) | 88.4% | [OpenAI — Introducing GPT-5](https://openai.com/index/introducing-gpt-5/) | 2025-08 | | HealthBench Hard | 46.2% | [OpenAI — Introducing GPT-5](https://openai.com/index/introducing-gpt-5/) | 2025-08 | *Scores above are reported by OpenAI and shown for context. They are not Benchgen measurements. SWE-bench runs used a fixed 477-task verified subset; the GPQA figure is for the GPT-5 pro variant.* ## GPT-5 vs Alternatives | Model | Context | SWE-bench Verified | Price (in/out per 1M) | |---|---|---|---| | **GPT-5** | 400K | 74.9% | $1.25 / $10 | | [Claude Opus 4.6](/models/anthropic/claude-opus-4-6) | 1M (beta) | 81.42% | $5 / $25 | | [Gemini 3 Pro](/models/google/gemini-3-pro) | 1M | 76.2% | See provider | For a builder, the trade-off is clear: GPT-5 is by far the cheapest of the three and fast by design, but its 400K context and SWE-bench score trail Claude Opus 4.6's larger window and higher coding result. The choice comes down to whether a workload needs the extra context and top-end coding reliability enough to pay several times more per token. *(Rival scores attributed to [Anthropic](https://www.anthropic.com/news/claude-opus-4-6) and [Google](https://blog.google/products/gemini/gemini-3/); not Benchgen measurements.)* ## How GPT-5 Performs on Real Agent Tasks GPT-5's headline numbers describe a model that is strong across coding, math, and science, and OpenAI emphasises gains specifically in instruction following and agentic tool use — the capabilities that let it carry out multi-step requests and coordinate across tools. The 74.9% on SWE-bench Verified is a genuine real-world coding signal rather than a saturated multiple-choice score, and OpenAI reports GPT-5 (thinking) matching its prior o3 model with 50–80% fewer output tokens. What those benchmarks cannot tell you is how GPT-5 behaves on a *specific* domain workflow, run many times, under a consistent rubric. A leaderboard number measures general capability; what agent builders actually need is reliable, repeated completion of the task in front of them — including how often the model fails, and why. That reliability-under-repetition view is what Benchgen is built to surface, and it is the dimension public benchmarks leave open. ## Use GPT-5 via API ```python from openai import OpenAI client = OpenAI(api_key="YOUR_API_KEY") response = client.chat.completions.create( model="gpt-5-2025-08-07", messages=[{"role": "user", "content": "Summarize this contract clause..."}], ) print(response.choices[0].message.content) ``` ## Related Models & Benchmarks - Compare: [Claude Opus 4.6](/models/anthropic/claude-opus-4-6) · [Gemini 3 Pro](/models/google/gemini-3-pro) · [Grok 4](/models/xai/grok-4) - Benchmarks: [SWE-bench Verified](/benchmarks/swe-bench-verified) · [GPQA Diamond](/benchmarks/gpqa-diamond) · [AIME 2025](/benchmarks/aime-2025) ## Frequently Asked Questions

What is GPT-5?

GPT-5 is OpenAI's flagship model, released 7 August 2025 — a unified system combining a fast model, a "GPT-5 thinking" reasoning model, and a router that selects between them automatically.

What is GPT-5's context window?

400,000 tokens — roughly 600 pages of text — with a maximum output of 128,000 tokens.

How much does GPT-5 cost?

$1.25 per million input tokens and $10 per million output tokens via the OpenAI API; cached input is $0.125 per million, and the Batch API halves both rates.

Is GPT-5 open source?

No. GPT-5 is a proprietary, closed model available only through OpenAI's API and ChatGPT; there are no downloadable weights.

What is GPT-5's knowledge cutoff?

September 30, 2024.

--- *Specs and scores sourced from [OpenAI's GPT-5 announcement](https://openai.com/index/introducing-gpt-5/) and [GPT-5 model docs](https://developers.openai.com/api/docs/models/gpt-5). Third-party benchmark scores attributed inline. Pricing cited to OpenAI. Last updated 2026-06-17.*

GPT-5 — Files