Skip to content
AI.info

The Pulse

OpenAI releases GPT-6.2 Astra and Sol for paid users and developers

GPT-6.2 Astra and Sol are available in ChatGPT Work, Codex and the OpenAI API. OpenAI says both models face tighter restrictions on higher-risk cybersecurity assistance after its capability assessments.

OpenAI releases GPT-6.2 Astra and Sol for paid users and developers

GPT-6.2 Astra and Sol are rolling out from today to Plus, Pro, Enterprise and Edu users in ChatGPT Work and Codex, with access also offered through OpenAI’s API; availability depends on plan, workspace permissions and rollout progress. OpenAI released the models on November 17 alongside a system card detailing its capability and safety assessments.

The rollout does not add either model to regular Chat. Astra is intended for more demanding projects, while Sol is positioned for everyday work.

Access and pricing

Pro 500 subscribers can use the Ultrafast option with either model. Pro 200 subscribers can use Ultrafast with Sol; the option is also available through the API.

The API model identifiers are gpt-6.2-astra and gpt-6.2-sol. Astra costs $8.00 per million input tokens and $40.00 per million output tokens. Sol costs $2.00 per million input tokens and $10.00 per million output tokens.

OpenAI describes Astra as a model for difficult projects involving work such as slideshows, websites and 3D designs. It describes Sol as a model for multistep agentic workflows, writing and collaboration.

Reported performance

On Agents’ Last Exam, Astra at the xhigh setting scored 5.2 points above GPT-6 Astra at max, at less than half the cost per task, according to OpenAI’s developer-community post. On FrontierCode 2.0, the company says Sol nearly matched Astra on merge-ready changes to real codebases at about one-fifth the cost.

The company also says Astra achieved new highs in its evaluations of professional work and scientific research at less than a third of GPT-6 Astra’s cost per task. Those are the company’s reported results rather than an independent assessment of performance across all uses.

Capability thresholds and safeguards

In its preparedness assessment, OpenAI determined that both models meet its Critical threshold for cybersecurity capabilities and its High threshold for biological and chemical capabilities. Astra also meets the High threshold for AI self-improvement, while Sol falls below it. The company says Astra is its first externally released model to meet that AI self-improvement threshold.

OpenAI tightened default restrictions on higher-risk cybersecurity assistance and requires additional vetting for more permissive access. The company is working to bring the models to Daybreak so approved defenders can use them with reduced refusals.

In simulated realistic workplace tasks, the models took improper actions about half as often as their respective predecessors, according to the system card. The card also reports regressions on model-only evaluations measuring whether models disclose obstacles encountered in their work; the Codex developer prompt substantially reduces those omissions across the evaluations that test for them.

After OpenAI adjusted the models’ refusal boundaries for high-risk requests, both improved on follow-up static jailbreak tests. The company says both still regressed slightly compared with GPT-6 Astra on high-risk biology prompts in its static jailbreak evaluations.

For its tracked-capability evaluations, OpenAI tested the models without production safeguards, using a variety of prompting and other methods where relevant. The company cautions that those results represent a lower bound on potential capabilities, because other approaches could elicit behavior not observed in the tests.

Sources

  1. OpenAI
  2. OpenAI Developer Community

Companies in this story

Related stories

Explore

More articles