OpenAI today announced GPT-5, its most capable model to date, in a livestream from San Francisco. The new system unifies what were previously separate reasoning, vision, and code models into a single API endpoint, and dramatically extends usable context to 1 million tokens.
What’s new
- Native multimodal input across text, images, audio and video frames
- Persistent memory that carries across sessions, opt-in per workspace
- First-class agentic tool-use (browser, code, files) at the model layer
- New `reasoning_effort` parameter that trades latency for depth
On internal benchmarks, OpenAI claims a 34% reduction in hallucination rate compared to GPT-4o on long-form factual tasks, and substantial improvements on SWE-bench Verified.
Why it matters
GPT-5 collapses the toolchain. Developers who previously stitched together GPT-4o, o1, and Whisper now get one endpoint that decides, internally, how much to "think" before answering. That alone should shorten many AI application stacks by hundreds of lines of orchestration code.
GPT-5 is the first model where we feel comfortable putting it directly in front of users without a safety harness.
— Sam Altman, CEO of OpenAI
Developer access
The API is rolling out today to existing paid customers. ChatGPT Plus, Team and Enterprise users get access in the chat product within 48 hours.
from openai import OpenAI
client = OpenAI()
resp = client.responses.create(
model="gpt-5",
input="Plan a 7-day Tokyo itinerary for a vegetarian foodie.",
reasoning_effort="medium"
)
print(resp.output_text)Early access partners include Stripe, Shopify, Figma and Notion, all of whom have updated their AI features today.