More Leaks Detail Some Of The New AI’s Main Features

More Leaks Detail Some Of The New AI’s Main Features

Menlo Ventures AI investor Deedy Das posted the following tweet on X, claiming OpenAI will launch GPT-5 next week. According to information from an anonymous source that provided “reasonable proof,” GPT-5 will support context windows of up to 1 million tokens (input) and up to 100,000 tokens (output). GPT-5 will also support the Model Context Protocol (MCP) standard, which lets AI models connect to external data sources and “parallel tool calls.” Das also said GPT-5 will support dynamic (short and long) reasoning, Code Interpreter, and other tools.

The leaker also mentioned the codenames OpenAI is supposedly using for GPT-5 models: “o3-alpha > nectarine (GPT-5) > lobster (mini) > starfish (nano).” Put differently, Nectarine, Lobster, and Starfish are the three GPT-5 models that will supposedly be available to users.

These codenames also appeared in a report from The Neuron earlier this week, which speculated that GPT-5 is being tested. “At least six anonymous models—Zenith, Summit, Lobster, Nectarine, Starfish, and o3-alpha—that are supposedly outperforming nearly every other known model,” have been spotted on LMArena. This report indicates Zenith might be the top-tier version of GPT-5, not Nectarine.

Separately, X user chatgpt21 posted screenshots supposedly coming from the head of design at Cursor, who has been testing GPT-5 Alpha. The images in the following tweet show an interaction with an AI model in a Mac app, probably Cursor. The model name is blurred but in a way that’s still legible. It reads gpt-5-alpha, likely a test AI model that some people have been able to use for a while. Finally, some testers have noticed a GPT-5-powered Smart Mode that Microsoft is preparing for Copilot.

Leave a Comment

Your email address will not be published. Required fields are marked *

Scroll to Top