What is OpenAI GPT-5?
OpenAI launched GPT-5 on August 7, 2025 as its flagship model system and the default model for ChatGPT users. At launch, it replaced the previous model picker in ChatGPT with a single system that automatically routed a request to a faster response mode or to a deeper "thinking" reasoning mode depending on the complexity of the task. Through the OpenAI API, developers selected GPT-5 explicitly rather than receiving it as an automatic default.
The launch family included several sized variants — gpt-5, gpt-5-mini, and gpt-5-nano — that traded latency and cost against reasoning depth. OpenAI introduced a higher-effort gpt-5-pro variant shortly afterward through the Responses API. OpenAI positioned GPT-5 as a substantial improvement over GPT-4o and GPT-4.5 in coding, multi-step reasoning, instruction following, and factual reliability, while also reporting a lower hallucination rate than prior models.
This page describes that launch-era generation rather than OpenAI's current flagship lineup. The dated API snapshot gpt-5-2025-08-07 is scheduled to shut down on December 11, 2026, and newer model generations have superseded it. For current model availability and selection guidance, consult OpenAI's model catalog.
Model Card
| Attribute | Details |
|---|
| Organization | OpenAI |
| Model name | GPT-5 |
| Model type | Unified multimodal model with routed reasoning |
| Release date (public) | August 7, 2025 |
| Availability at launch | ChatGPT (default model), OpenAI API (explicitly selected) |
| Context window (API) | Up to 400K tokens total (272K input / 128K output) |
| Launch-era variants | gpt-5, gpt-5-mini, gpt-5-nano, gpt-5-pro (introduced later) |
| Parameter count | Not publicly disclosed |
| Training compute/cost | Not publicly disclosed |
OpenAI has not published parameter counts, training compute, or training cost figures for GPT-5; any specific numbers circulating for these should be treated as unverified.
Historical note: pre-release speculation (2023–2024)
Before GPT-5 was announced, its development was the subject of extensive public speculation. In 2023 and early 2024, Sam Altman and other OpenAI staff referred to internal reasoning-focused efforts under the code names Q* and Strawberry, and to a next-generation flagship model under the code name Orion. At various points, Altman said training had not yet started (April 2023), then confirmed work was underway (November 2023), while declining to give a release date. Commentators speculated about parameter counts in the trillions, training runs using hundreds of thousands of H100 GPUs, and costs in the billions of dollars, alongside a rumored late-2024 release.
Archived pre-release speculation snapshot
| Attribute | Details |
|---|
| Codename | Q* (Strawberry) / Orion / Gobi / Arrakis (GPT-4.5) |
| Model type | Frontier Multimodal Reasoning Model |
| Parameter count | 2T-5T (2,000B-5,000B) |
| Scaling Principle | 10-20x GPT-4 Compute, 10x RLHF/AIF Reasoning Data |
| Training Hardware | 250-500k H100s |
| Dataset size (tokens) | 40T-100T (80TB-200TB) |
| Training data end date | Dec/2023 |
| Training start date | Jan25/2024 |
| Total training time | 3 Months |
| Training cost | 1.25-2.5 Billion USD |
| Release date (public) | Nov/2024 |
| Paper | TBA |
| Playground | TBA |
None of these figures were ever confirmed by OpenAI, and the actual GPT-5 released in August 2025 did not match the specific specifications that were speculated about at the time. This section, images, and table are preserved for historical context only and should not be read as a description of the shipped model.
