Last updated:
GPT
OpenAI's flagship model family — the engine behind ChatGPT, now the GPT-6 generation (Astra, Sol, Luna and Instant).
Model Facts
Latest model: GPT-6 Instant
- Provider
- OpenAI
- Released
- Sep 4, 2026
- Context window
- 1M tokens (API)
- Weights
- Closed
Benchmarks
96.3%
GPQA Diamond
GPT-6 Astra
Full Model List
| Model | Released | Note |
|---|---|---|
| GPT-6 Instant | Oct 8, 2026 | fast default model with interactive "Intelligent UI" widgets |
| GPT-6 Sol / Luna | Sep 22, 2026 | tiered GPT-6 variants, rolled out to all ChatGPT users in October |
| GPT-6 Astra | Sep 4, 2026 | first GPT-6 model, trained on 100,000+ GPUs at the Stargate site in Texas |
| GPT-5.5 | Apr 23, 2026 | previous flagship, 1M-token API context |
| GPT-5 | Aug 7, 2025 | unified reasoning + non-reasoning model |
| GPT-4o | May 13, 2024 | natively multimodal flagship |
| GPT-4 | Mar 14, 2023 | first GPT-4 generation |
| GPT-3.5 | Nov 30, 2022 | original ChatGPT launch |
- Official model page:
- openai.com
- Official documentation:
- platform.openai.com/docs
- Other sources:
- GPT-6 (Wikipedia)
- OpenAI deployment safety hub
What it is
GPT-6 is OpenAI's current model generation and the engine behind ChatGPT and the OpenAI API. The flagship, GPT-6 Astra, reached the public on September 4, 2026 after a one-day limited preview. OpenAI describes it as faster and able to do more than any earlier iteration, with particular strength in coding, math, and operating computers and web browsers.
The GPT-6 lineup
- Astra — the top-capability model, with a roughly 1M-token context window and up to 128K output tokens in the API.
- Sol and Luna — tiered variants released on September 22, 2026 and rolled out to everyone in ChatGPT (free and paid) in October, replacing the GPT-5.6 versions.
- Instant — a fast model released in early October 2026 with an "Intelligent UI" that answers with interactive widgets.
Where these numbers come from
The GPQA Diamond figure is the one most widely reported for GPT-6 Astra; independent evaluators that run their own harness report lower GPQA scores, so treat it as a vendor-style number rather than a neutral comparison. OpenAI has published safety and capability evaluations for Sol and Luna on its deployment-safety hub.
Notable traits
- Ships in multiple tiers by cost and speed (Astra / Sol / Luna / Instant) rather than one model.
- Training reportedly uses a "recurrent depth" technique, which safety researchers say can make the model's reasoning harder to monitor.
- Closed weights — available only through OpenAI's API and products, not for local download.