The News, Explained

OpenAI launched GPT-6.2 Astra and Sol on November 17. They’re available in ChatGPT Work and Codex to Plus, Pro, Enterprise and Edu users. Eligible Enterprise and Edu users need their workspace admin to enable access. Pro 500 subscribers can use the faster Ultrafast option with either model; Pro 200 subscribers get that option starting with Sol. The requirements for regular model access and the faster option are separate. Source

OpenAI positions Astra for complex work such as drawing on several sources to prepare documents or slides, or making substantial changes to existing software. It presents Sol as a model for everyday work and coding. If you’re updating a team presentation, for example, you could give Astra the existing template and the new source files, then check how closely it follows the format. For a first draft of one document or a small code review, Sol is a reasonable place to start. Which model gets there with fewer edits will depend on the material and the task. Source

Developers using the API separately pay standard rates by tokens, the chunks of text the model reads and writes. Per million tokens, Astra costs $8 for regular input, $0.40 for reused input served from cache, and $40 for output. Sol costs $2, $0.10 and $10 respectively, one quarter of Astra’s rate in each case. That comparison assumes the same number and mix of tokens. The actual cost of a job also depends on how much each model uses and whether it needs another attempt. Source Source

OYOPICK’s Take

When choosing a model, look at what it was able to verify alongside the finished file. OpenAI says both new models were more likely than GPT-6 Astra to leave out blockers such as a missing file or an unavailable search tool in two evaluations. It says changes to Codex instructions reduced those omissions. If you’re testing the models for work, give both the same materials and compare the cost and number of revisions. Asking each to list missing sources and checks it couldn’t run helps you judge how much review time the result still needs. Source