GPT-6 in ChatGPT: Intelligent UI and What Developers Get

On October 7 OpenAI made GPT-6 the default model in ChatGPT for paid plans, and free users got it a day later. It came with Intelligent UI, which lets ChatGPT answer with buttons, forms, charts and small working tools instead of only text. The names are confusing, because there are now four GPT-6 models released over five weeks. This post sorts out which model is which, what each costs in the API, and what you as a developer can and cannot use yet.

The GPT-6 family in one table#

The September releases went to ChatGPT's Work mode, to Codex and to the API. The ordinary Chat tab kept GPT-5.6 until this week. On October 7 and 8, OpenAI shipped new versions of Sol and Luna "tuned for everyday conversation" to the Chat tab: Sol for Plus, Pro, Business and Enterprise, Luna for Free and Go. The release notes say Work and Codex keep the September versions, and the Pro reasoning option stays on Astra, without Intelligent UI.

What Intelligent UI does#

Instead of a wall of text, GPT-6 now picks the format for each answer. According to OpenAI's announcement, it can combine text with graphics, tappable buttons, forms, charts and interactive widgets, and it still answers in plain text when that works best. OpenAI's examples include roast timings next to a recipe, road-trip stops on a map, side-by-side comparisons, diagrams you change by adjusting inputs, and tools built on request, such as a savings calculator or a bill splitter.

Under the hood, OpenAI describes "native, streamable components" and a compiler that renders the interface progressively while the answer streams. That is the interesting engineering part: the model emits structured components instead of HTML, and the client decides how to draw them. It is the same idea as server-driven UI in mobile apps, now generated per answer.

What developers can use today#

All four models are in the API. These are the standard prices per million tokens from OpenAI's pricing page, for prompts up to 272,000 tokens:

Above 272,000 input tokens the whole request costs twice as much for input and one and a half times as much for output. Batch and Flex processing are half price. Astra, GPT-6.1 Sol and Luna accept up to about 1.05 million tokens of context, with up to 128,000 tokens of output.

A call through the Responses API looks the same as for earlier models:

from openai import OpenAI

client = OpenAI()  # reads OPENAI_API_KEY

response = client.responses.create(
    model="gpt-6.1-sol",
    input="Write a Django REST Framework serializer for a Book with title, isbn and author.",
)
print(response.output_text)

Things that trip people up#

• Astra is stricter. It supports tool calling only through the Responses API, and it does not accept temperature or top_p. Remove them when you switch from an older model.
• chat-latest moves. Since October 7 it points at the newest ChatGPT model, so its behavior can change under you. Use a fixed model name in production.
• Codex recommends GPT-6.1 Sol as its default for coding. Codex can also run Astra, which keeps notes across context windows on long tasks.

How to choose a model#

• High-volume, simple work such as classification, extraction or short summaries: start with Luna. It costs a hundredth of Astra per token.
• Coding and agents: GPT-6.1 Sol. OpenAI says it nearly matches Astra on agentic coding at a fifth of Astra's price.
• The hardest reasoning, where a wrong answer costs more than the tokens: Astra, and measure whether it is really better on your own tasks.

On public coding benchmarks, OpenAI's own table for Astra puts it ahead on some and behind Anthropic's Claude Opus 5 on others. Independent results differ from vendor tables. The coding agent trends post has the current Terminal-Bench leaderboard, where agents are tested end to end.

Safety ratings and a difficult week#

OpenAI rates the October Sol and Luna as High capability in cybersecurity, below its Critical level, and High in biology and chemistry, according to its deployment safety page. Astra and GPT-6.1 Sol are rated Critical in cybersecurity. The safeguards include real-time monitors that block or escalate risky requests, identity checks for advanced cyber and biology access, and new teen-safety classifiers.

The launch also came during a hard week for OpenAI's safety work. The Wall Street Journal, CNN and AP reported that OpenAI dismissed three safety researchers, who dispute the reasons given. OpenAI says the dismissals were about policies on handling sensitive company information, "not about safety concerns or speaking out". Separately, a staff member who describes leading launch safety reports resigned with a public essay criticizing the company's safety culture. None of this changes the API today, but it is worth following if you build on these models.

Frequently asked questions#