Model Drop

GPT-5

August 7, 2025 2 min read

GPT-5, which OpenAI shipped on August 7, 2025, is a unified system that routes each request between a fast model and a deeper thinking model, and it became the default in ChatGPT for every signed-in user.

Model: GPT-5 (gpt-5)
Maker: OpenAI
Ship date: August 7, 2025
Pricing: $1.25 / $10 per M tokens
Context: 272K input tokens (400K total with output)
Where you can use it:
- ChatGPT on Free, Plus, Pro and Team
- The API in three sizes, plus Codex CLI
When to reach for it:
- Agentic coding and long tool-call chains
- Everyday ChatGPT work, as the default
- Skip high effort when speed matters most
How it compares:
- Same list price as Gemini 2.5 Pro
- Far cheaper per token than Claude Opus 4.1
What makes it different:
- A router picks fast or thinking mode
- New minimal effort and verbosity controls
Early reactions: Launch-day router failures and the loss of GPT-4o drew complaints, and Sam Altman said OpenAI was looking at letting Plus users keep 4o.
Read more: Introducing GPT-5, Introducing GPT-5 for developers, GPT-5 System Card

The benchmarks

Scores are GPT-5 at high reasoning effort against OpenAI's own o3 (high) and GPT-4.1, plus ChatGPT agent on BrowseComp; OpenAI showed no outside rivals, and its SWE-bench score skips 23 of 500 tasks.

OpenAI's numbers

SWE-bench Verified (real GitHub fixes): 74.9%. +5.8 over o3 (69.1%), +20.3 over GPT-4.1 (54.6%).
Aider Polyglot (diff-format code edits): 88.0%. +8.4 over o3 (79.6%), +35.1 over GPT-4.1 (52.9%).
AIME 2025 (competition math, no tools): 94.6%. +5.7 over o3 (88.9%), +48.2 over GPT-4.1 (46.4%).
GPQA Diamond (PhD-level science, no tools): 85.7%. +2.4 over o3 (83.3%), +19.4 over GPT-4.1 (66.3%).
Humanity's Last Exam (no tools) (expert-level questions): 24.8%. +4.6 over o3 (20.2%), +19.4 over GPT-4.1 (5.4%).
τ²-bench Telecom (customer-support tool use): 96.7%. +38.5 over o3 (58.2%), +62.7 over GPT-4.1 (34.0%).
τ²-bench Airline (airline-support tool use): 62.6%. −2.2 under o3 (64.8%), +6.6 over GPT-4.1 (56.0%).
BrowseComp (hard web research): 54.9%. +5.2 over o3 (49.7%), −14.0 under ChatGPT agent (68.9%).

Independent numbers

Artificial Analysis Intelligence Index (current v4.3.2, AA estimate): 23.0. +0.2 over Claude Opus 4.1 (22.8), +0.5 over Grok 4 (22.5), +6.9 over Gemini 2.5 Pro (16.1).
ARC-AGI-1 (abstract visual puzzles): 65.7%. +4.9 over o3 (high) (60.8%), −1.0 under Grok 4 (66.7%).
ARC-AGI-2 (harder visual puzzles): 9.9%. +3.4 over o3 (high) (6.5%), −6.1 under Grok 4 (16.0%).