GPT-6 comes to free ChatGPT, and the answers now come with buttons
OpenAI moved ChatGPT's chat tab to GPT-6 and added Intelligent UI, which answers with charts, forms, and small tools. Here is how the GPT-6 family benchmarks.
- Model
- GPT-6 (Astra, Sol, Luna)
- Maker
- OpenAI
- Context
- 1M (Astra)
- License
- Proprietary
- Weights
- Closed
GPT-6 has been out for paying customers since September. On October 7, OpenAI started moving everyone else onto it, and changed what a ChatGPT answer can look like while it was at it.
According to Let’s Data Science, Plus, Pro, Business, and Enterprise plans get GPT-6 Sol in the chat tab, and Free and Go users were due to get the smaller GPT-6 Luna from October 8. The models behind Work and Codex are unchanged.
Intelligent UI
The more interesting half of the update is a response format OpenAI calls Intelligent UI. Instead of always replying in text, ChatGPT can answer with interactive pieces: charts, buttons, forms, and small tools built for the question in front of it. Comparisons, diagrams, maps, and calculators are the examples given.
The format is chosen per answer. If plain text is the clearest reply, it stays plain text, and users are reported to be getting a setting to turn the visuals down.
For builders, this is the part to watch. A chat answer that is a working calculator rather than a description of one moves ChatGPT closer to being a place where small apps get generated on demand. We could not load OpenAI’s own page to check the technical details, so we are not going to describe how it works under the hood.
The benchmarks
OpenAI did not publish new scores with this rollout in anything we could read. The numbers that exist are for the family’s flagship, GPT-6 Astra, released on September 3. These are OpenAI’s figures as collected by Vellum, with GPT-5.6 Sol as the comparison.
| Benchmark | GPT-6 Astra | GPT-5.6 Sol |
|---|---|---|
| Terminal-Bench 4.0 | 57.7 | 37.3 |
| DeepSWE v1.1 | 74.1% | 72.7% |
| OSWorld 2.0 | 72.6% | 65.7% |
| ScreenSpot-Pro | 92.7% | 76.9% |
| AutomationBench | 41.4 | 18.1 |
| BrowseComp | 91.5% | 90.4% |
| FrontierMath Tier 4 | 97.6 | n/a |
| Humanity’s Last Exam, with tools | 57.2% | n/a |
Vellum lists several caveats worth repeating. Every score was run at maximum effort, which flatters results and hides cost. OpenAI helped fund FrontierMath and has exclusive access to part of it. And on the one independent composite in the article, the Artificial Analysis Intelligence Index, Astra scores 61.2, behind Anthropic’s Fable 5.1, Opus 5, and Fable 5 at 62.1 to 65.7.
Astra is also not what most people will be talking to. It costs $10 per million input tokens and $50 per million output tokens.
What free users actually get
For Luna, the small model now behind free ChatGPT, the best public numbers come from a competitor. Anthropic’s launch of Claude Haiku 5.5 included Luna in its tables: 48.9% on OSWorld 2.1, 16.4% on Terminal-Bench 4.0, 42.4% on FrontierCode 1.1 Main. Treat a rival’s measurements with the usual care.
Independent testing is kinder. The Decoder reports that Artificial Analysis scores Luna at 38 on its Intelligence Index, a few points behind the best small models but using far fewer tokens per task, which makes it the more cost-efficient choice in that analysis.
Try it yourself
Open ChatGPT. If you ask for a comparison or a calculation and get something you can click, you have it. We have not tested Intelligent UI ourselves: this is based on Let’s Data Science’s and Vellum’s reporting, and OpenAI’s announcement page would not load for us.