Robots Atlas>ROBOTS ATLAS
GPT-5 Thinking

GPT-5 Thinking

5 · Family: GPT
GPT-5 variant with deep reasoning mode, available in ChatGPT and via the API through the reasoning_effort parameter. Released August 7, 2025.
✓ Active✓ Public accessReasoning modelMultimodalLLM📁 GPT
Context window
400K
tokens
Max output
128,000
tokens
Release date
7 August 2025
Access:APIHostedDeployment:☁ Cloud

Overview

GPT-5 Thinking is the deep reasoning variant within the unified GPT-5 system released by OpenAI on August 7, 2025. The GPT-5 system consists of a fast model answering most queries, a deeper reasoning model (GPT-5 Thinking), and a real-time router that selects the path based on conversation type, complexity, tool needs, and user intent. In ChatGPT, GPT-5 Thinking is explicitly selectable from the model picker for paid users, or can be triggered by phrases such as "think hard about this". In the API, the same GPT-5 model (identifier gpt-5, snapshot gpt-5-2025-08-07) exposes a reasoning_effort parameter with values minimal/low/medium/high that controls the depth of the chain-of-thought. The context window is 400,000 tokens, maximum output is 128,000 tokens. Knowledge cutoff: September 30, 2024. Input modalities are text and image, output is text. API pricing: USD 1.25 per 1M input tokens (USD 0.125 cached) and USD 10 per 1M output tokens. The model was trained on Microsoft Azure AI supercomputers.

Classification
Reasoning modelMultimodalLLM
Family: GPT
Access & deployment
APIHosted
Cloud
Weights: Closed
Key parameters
📏 Context: 400K
Tools
📥 Input: text, image

Technical specification

Context window
400K
tokens
Max output tokens
128,000
tokens per response
Knowledge cutoff
30 Sept 2024
Knowledge boundary
Features:Tool use
Modalities
⬇ Input
textimage
⬆ Output
textcode

Capabilities and applications

Native model capabilities
Reasoning
Category: reasoning
Multi-step reasoning
Category: reasoning
Coding
Category: coding
Long context
Category: reasoning
Multilingual
Category: language
Image understanding
Category: vision
Function Calling
Category: planning
Planning
Category: planning
Parallel Tool Calls
Ability to invoke multiple external tools simultaneously while generating a response.
Category: reasoning
Agentic capability
The model's ability to autonomously plan and execute multi-step tasks by sequentially using tools, maintaining context, and adapting to intermediate results.
Category: planning
Structured output
Category: structured_generation
Computer use
The model's ability to operate a computer interface by interpreting screenshots and generating actions such as clicks, typing, and navigating applications.
Category: planning

Benchmark results

6 benchmarks
SWE-bench
accuracy · SWE-bench Verified, fixed subset n=477, high reasoning effort
74.9%
📅 7 Aug 2025📄 OpenAI announcement (Introducing GPT-5)
MMMU
accuracy · Multimodal understanding, high reasoning effort
84.2%
📅 7 Aug 2025📄 OpenAI announcement (Introducing GPT-5)
GPQA
accuracy · GPQA Diamond, GPT-5 Thinking without tools (GPT-5 Pro reaches 88.4%)
85.7%
📅 7 Aug 2025📄 OpenAI announcement (Introducing GPT-5)
AIME 2025
accuracy · Without tools, high reasoning effort
94.6%
📅 7 Aug 2025📄 OpenAI announcement (Introducing GPT-5)
Aider Polyglot
accuracy
88%
📅 7 Aug 2025📄 OpenAI announcement (Introducing GPT-5)
HealthBench Hard
accuracy
46.2%
📅 7 Aug 2025📄 OpenAI announcement (Introducing GPT-5)

Pricing

Technical architecture

Deployment and security