OpenAI GPT-6.2 Astra and Sol Launch for Paid Users
OpenAI has launched GPT-6.2 Astra and Sol. The announcement lists Plus, Pro, Enterprise, and Edu users as eligible for Work and Codex, but does not specify when activation will be complete for each account in Korea.
According to OpenAI's announcement, GPT-6.2 Astra and Sol are available in Work and Codex to Plus, Pro, Enterprise, and Edu users.
The Enterprise and Edu help documentation states that both models are disabled by default in the Work and Codex desktop previews.
OpenAI's announcement lists Astra's standard API prices per 1 million tokens as $8 for input, $0.40 for cached input, and $40 for output.
OpenAI's announcement lists Sol's API prices per 1 million tokens as $2 for input, $0.10 for cached input, and $10 for output.
The context window specified in the Astra model documentation is 1,050,000 tokens.
OpenAI launched GPT-6.2 Astra and Sol on November 17, 2026, the day of the announcement. The announcement lists Plus, Pro, Enterprise, and Edu users as eligible. The models are available in ChatGPT Work and Codex.
The information in this article was checked as of November 19, 2026, based on OpenAI's announcement, model documentation, and help articles.
Astra and Sol launch and Responses API changes
OpenAI released GPT-6.2 Astra and GPT-6.2 Sol for external use. According to the Sol model documentation, inputs are text and images. The output format is text.
The Using GPT-6.2 documentation explains what happens when conversation history is edited. Starting with GPT-6.2, the Responses API drops reasoning items if earlier conversation history is edited. The documentation calls these changes prefix edits.
Available in Work and Codex for Plus, Pro, Enterprise, and Edu
According to the OpenAI announcement, both models are available on the following plans from the announcement date. They can be accessed through ChatGPT Work and Codex. However, they are disabled by default in the Enterprise and Edu desktop previews.
Plan | Access announced | Access condition to check
Plus | ChatGPT Work, Codex | Model picker in each service
Pro | ChatGPT Work, Codex | Model picker in each service
Enterprise | ChatGPT Work, Codex | Whether enabled in the desktop preview
Edu | ChatGPT Work, Codex | Whether enabled in the desktop preview
The model selection help article explains the model pickers in Work and Codex. The existing model options in regular Chat remain unchanged. Access to these models should therefore not be judged solely by the picker in regular Chat.
The Enterprise and Edu help article specifies separate conditions. Both models are disabled by default in the Work and Codex desktop previews. Even on an eligible plan, workspace access settings must be checked before use.
API pricing and Astra's input and output limits
According to OpenAI's Korean-language announcement, API rates for the two models are as follows. The amounts in the table are based on the November 17, 2026 announcement. All rates are in US dollars per 1 million tokens of the corresponding type.
Model | Input | Cached input | Output
GPT-6.2 Astra | $8 | $0.40 | $40
GPT-6.2 Sol | $2 | $0.10 | $10
OpenAI described Astra's rates as standard API pricing. It said Sol retains the pricing of GPT-6.1 Sol. This table lists API rates, not ChatGPT subscription fees.
The Astra model documentation specifies a context window of 1,050,000 tokens. Maximum input and maximum output are listed separately. The full context window should therefore not be read as the input limit.
Item in the Astra documentation | Limit
Context window | 1,050,000 tokens
Maximum input | 922,000 tokens
Maximum output | 128,000 tokens
OpenAI's own evaluations and AA-3D measurements by a separate organization
OpenAI's capability assessments should be distinguished as the company's own claims. The GPT-6.2 System Card evaluated the capabilities of both models. OpenAI said they are the most capable models it has deployed externally.
AA-3D scores published by Artificial Analysis
Artificial Analysis said Astra leads AA-3D in its separate measurements. The scores presented in its November 17, 2026 analysis are below. These figures are not OpenAI's own measurements.
Model | AA-3D score | Measured and published by
GPT-6.2 Astra | 1318 points | Artificial Analysis
Claude Fable 5.2 | 1271 points | Artificial Analysis
GPT-6.2 Sol | 1259 points | Artificial Analysis
The comparison in this table is limited to AA-3D scores. These figures do not indicate the quality of Korean-language responses. These results alone cannot establish performance rankings across all tasks.
Activation timing in Korea and Business eligibility were not specified
The announcement did not confirm when activation would actually be completed for individual accounts in Korea. It gives an announcement date but does not specify the exact release time or time zone. A launch time converted to Korean time therefore cannot be stated with certainty.
· When activation will be completed for individual accounts in Korea: not specified.
· Eligibility to use these models on the Business plan: not specified.
· Schedule for retiring existing models or automatically switching users: not specified.
· API access conditions for each model in Korea: cannot be confirmed from the official wording cited.
Check the model picker and access settings before use in Korea
To use the models in Korea, check both your plan and whether you actually have model access. Before deciding to pay, compare the eligibility stated in the announcement with what appears in your account. If integrating the API, also check input and output limits and how conversation history is handled.
· Check whether your current plan is listed as eligible in the announcement. Do not assume Business is eligible based on this announcement alone.
· Check the model picker in ChatGPT Work or Codex. Regular Chat retains its existing model options.
· If you use Enterprise or Edu, check your workspace access settings. The desktop preview is disabled by default.
· When reviewing API costs, distinguish between input, cached input, and output rates. Also check Astra's maximum input and maximum output separately.
· If your implementation edits earlier conversation history, check the Responses API's behavior. Starting with GPT-6.2, reasoning items are dropped after those edits.