What the GPT model family is
GPT is OpenAI’s family of general-purpose models; this page covers the models themselves and the developer interface. It sits at a different layer from ChatGPT: ChatGPT is the assistant product for individuals and organizations, while GPT is the model layer behind such products, also exposed to developers through the OpenAI API platform.
As of September 2026, the official model catalog lists the flagships GPT-6 Astra (gpt-6-astra) and GPT-5.6 Sol, GPT-5.6 Terra, and GPT-5.6 Luna, plus specialized image, realtime voice, and transcription models. Flagship models support reasoning and tools including function calling, web search, file search, and computer use. The GPT-6 Astra announcement describes it as the successor to GPT-5.6 Sol, rolling out first to a limited set of organizations and then to all ChatGPT Plus, Pro, Business, and Enterprise users over the following days, available through the OpenAI API, Microsoft Azure, and AWS Bedrock; Enterprise workspaces have access off by default at launch. Model names, availability, and pricing change, so check the official model catalog before relying on any of this.
Uses and boundaries
The same name covers two typical uses: end users pick a model inside ChatGPT for Q&A, writing, and analysis, while developers create an API key on the OpenAI API platform and integrate the models into their own applications and agent workflows. The full coding agent for software development is a separate product, Codex, which is likewise built on these models.
Do not infer capability from a model name: models in the family target different tasks and cost tiers and are updated on a rolling basis. ChatGPT subscriptions and API usage are billed separately — API usage is not included in a ChatGPT plan. Available features also depend on account type, region, and organization policy; for example, the announcement states that Enterprise workspaces have GPT-6 Astra access off by default and an admin must enable it.
Discussion in the show
Weekly #002 was recorded on the day GPT-6 Astra launched. In the chapters “GPT-6 Astra：从 3D 建模聊起” (GPT-6 Astra: starting from 3D modeling) and “Blender 与代码：给模型合适的工具” (Blender and code: giving the model the right tools), Guizang describes using Astra for 3D and modeling and gives his overall assessment: no obvious gap versus Fable in his own agent work, but a major opportunity in multimodal and 3D; he also observed GPT-6 modeling through Blender’s Python interface — writing code rather than clicking the UI, which kept token usage low. Juzi relays a new benchmark comparison saying 3D and Computer Use are where Astra clearly stands out. Yang Pan describes producing and publishing the first episode and the website entirely with Astra and relays news of OpenAI’s earlier acquisition of a computer-control team and a bulk Mac purchase, adding his own guess that it relates to training agents — that part is hearsay plus speculation. See the 3D modeling chapter, the Blender chapter, and the podcast-production chapter of the Chinese transcript.
The show also covered weaknesses. In the writing chapter, Yang Pan argues GPT is poor at writing, and Juzi adds that novel writing only became viable from GPT 5.5 onward while domestic models still could not write Chinese web novels. In the image-editing chapter, Yang Pan cites unchanged regions staying perfectly stable across successive edits as the strength of GPT Image 2.5 (the official GPT-Image-2.5 image series), while Guizang points to lingering artifacts and aesthetics issues. In the chapter on models “getting dumber,” Yang Pan argues the perception is more likely caused by bugs, noting that both Claude and GPT had waves of user complaints traced to bugs and fixed this year; the community “Juice value” claims Xiangyang Qiaomu relays remain unverified community lore. See the writing, image, and model-perception chapters of the Chinese transcript.
On the timeline: Weekly #001’s “Computer Use 正在怎样改变 Vibe Coding 的验收” (how Computer Use is changing vibe-coding acceptance) was recorded before GPT-6 launched, with Guizang relaying claims about its new architecture; by Weekly #002’s recording day, Astra had shipped. These are participant experiences, retellings, and predictions from the Chinese transcript — not official endorsements or independent evaluations.
Frequently asked questions
How is GPT different from ChatGPT?
GPT is OpenAI’s model family; ChatGPT is OpenAI’s assistant product built on those models. The same models can be used directly inside ChatGPT or called separately by developers: ChatGPT is billed to users through plans, while API usage is billed separately on the OpenAI API platform.
What is the newest GPT model?
As of September 2026, the newest model in the official catalog is GPT-6 Astra (gpt-6-astra). The announcement says it rolled out first to a limited set of organizations and then to ChatGPT paid plans over the following days; whether your account can select it depends on what ChatGPT and the catalog actually show you.
Which models make up the GPT-5.6 series?
The official model catalog lists three: GPT-5.6 Sol (flagship for complex professional work), GPT-5.6 Terra (balances intelligence and cost), and GPT-5.6 Luna (optimized for cost-sensitive workloads). Context limits and rates for each are as currently labeled in the catalog.
Where do I get a GPT API key?
Register on the OpenAI API platform and create an API key, then pick a model as described in the model documentation. API keys and ChatGPT subscriptions are separate; keep the key out of public pages and repositories.
How is the GPT API priced?
The API charges per input and output token, with rates varying by model; ChatGPT plans and the API are separate billing systems. Current per-model pricing is listed in the official model catalog.
Why do searches for “GPT” return unrelated results?
GPT has entirely different meanings in other fields (a medical test marker, a disk partition table format), so same-name content mixes into results. To find OpenAI’s models, add words like OpenAI, AI, or ChatGPT to the query, or go straight to openai.com.