Three months after GPT 6 Astra arrived, OpenAI already has a successor. On 17 November it announced GPT 6.2 Astra and GPT 6.2 Sol on its community forum, available from today in ChatGPT Work, Codex and the API.
The focus of this release is not just a smarter model but a cheaper one to run. OpenAI says GPT 6.2 Astra beats GPT 6 Astra while spending less than a third as much per task, and it has cut the price per token as well.
What GPT 6.2 Astra brings
Astra is OpenAI's most capable model, the one built for long tasks, complex code and agents that work on their own for hours. According to the company, on the Agents' Last Exam benchmark GPT 6.2 Astra at the xhigh reasoning setting scores 5.2 points higher than GPT 6 Astra at its maximum setting, at less than half the cost.
OpenAI also measures it against the competition, claiming it scores 5.5 points above Anthropic's Claude Fable 5.2 at 5% of the cost per task. That figure comes from OpenAI itself and there are no independent results yet. For where Anthropic's model stands today, see our comparison of Claude Fable 5.2 and Opus 5.5.
Sol nearly matches it for much less
GPT 6.2 Sol is the mid-tier model. On FrontierCode 2.0, a coding benchmark, it comes close to Astra at around a fifth of the cost. For a lot of everyday work, such as summaries, emails or straightforward code, it may be the more sensible choice.
Ultrafast mode, which answers much faster in exchange for using more of your usage limit, is available on both models in the Pro 500 plan. For the first time it is also coming to the Pro 200 plan, starting with Sol.
Stronger at cyber security, with tighter guardrails
OpenAI calls GPT 6.2 Astra its most capable model yet for cyber security work, with a 22-point gain on ExploitBench, a test of how well a model finds and exploits software flaws. That is why it ships with stricter safeguards around those kinds of requests.
The system card adds another first: it is the first OpenAI model rated High for AI self-improvement, meaning its ability to help build better AI models. Reaching that level requires OpenAI to apply extra safeguards before release.

GPT 6.2 pricing and who can use it
In ChatGPT, both models are available on the Plus, Pro, Enterprise and Edu plans. OpenAI has said nothing about the Free and Go plans, so for now they stay on the earlier models. Our guide to ChatGPT plans in Australia covers what each one includes.
In the API, GPT 6.2 Astra costs about $11.87 per million input tokens and $59.33 per million output tokens at the RBA exchange rate. That is 20% less than GPT 6 Astra, which cost about $14.83 and $74.16. GPT 6.2 Sol keeps its price of about $2.97 for input and $14.83 for output.
Why Australian businesses should care
If your business runs on the OpenAI API, the maths is simple: better quality for less money. But test before you switch. A new model can respond differently to the same prompts, and that can break workflows you have already automated. The careful approach is to compare results and costs on real tasks for a few days.
For teams using ChatGPT at work, the upgrade arrives on its own. What is worth reviewing is which model each team uses and for what: Sol may cover most tasks, leaving Astra for the hard ones. If you want help sorting out which models your business uses and with what data, that is what our AI consulting team does.
What comes next
The race between OpenAI, Anthropic and Google is no longer just about who ships the most powerful model, but who delivers it most cheaply per task. We will need independent testing to know whether GPT 6.2 Astra lives up to its claims against Claude Fable 5.2. In the meantime, our comparison of GPT 6 Astra, Sol and Luna helps you choose within the family.
