OpenAI Launches GPT-6.2, Cuts Astra Token Prices by 20% for Developers
Key Facts
OpenAI announced the launch of GPT-6.2 Astra and GPT-6.2 Sol on November 17, describing them as the most capable models it had made available externally to date. The company is making both models available to Plus, Pro, Enterprise and Edu subscribers in ChatGPT Work and Codex, and to developers through the OpenAI API. The Ultrafast option is available through the API from launch. OpenAI is positioning Astra for complex professional projects, while offering Sol for everyday work. The company said the models deliver better results on difficult tasks using fewer output tokens.
Astra’s standard prices per million tokens are $8 for input, $0.40 for cached input and $40 for output. The company says its input and output prices are 20% lower than those for GPT-6 Astra, while cached input is 60% cheaper. Sol retains GPT-6.1 Sol’s prices: $2 for input, $0.10 for cached input and $10 for output per million tokens. The pricing paths for the two models therefore differ at launch: Astra is cheaper than its predecessor, while Sol’s prices are unchanged. OpenAI said the reduction in tokens needed for difficult tasks comes alongside improved results in both models.
In its internal GDPval v2 test, OpenAI said Astra scored 6.2 percentage points higher than GPT-6 Astra while using 22% fewer output tokens. It said Sol improved by 5.3 points over GPT-6.1 Sol in the same test. In the Agents’ Last Exam test for multi-step professional tasks, the company said Astra scored 5.2 points ahead of GPT-6 Astra and 5.5 points ahead of Claude Fable 5.2. It estimated Astra’s cost in the comparison at less than one-third that of GPT-6 Astra and 5% of the cost of Claude Fable 5.2. OpenAI also said Astra scored 9 percentage points higher than GPT-6 Astra on the Terminal-Bench-Science 0.1 test, at less than one-third of its cost.
In other tests, OpenAI said Sol came within about 0.5 points of Astra on FrontierCode 2.0 at the highest reasoning level, at about one-fifth of Astra’s cost. It added that Astra handles presentations, spreadsheets and professional documents better, and follows user-provided templates and industry standards more closely. The company says gains on multi-step tasks are accompanied by lower output-token use. These test results and costs are comparisons announced by OpenAI for the models cited.
OpenAI framed the launch of the two models in the context of advanced-capability risks, classifying them at the Critical threshold for cyber capabilities and the High threshold for biological and chemical capabilities. Astra alone reached the High threshold for AI self-improvement capabilities, a classification the company says measures a model’s ability to accelerate AI research. OpenAI said both models were better than GPT-6 Astra and GPT-6.1 Sol at developing working exploits for known software vulnerabilities. The company said it had tightened default restrictions on high-risk cyber assistance and required additional screening to grant broader permissions. In simulated UK AISI evaluation tests, Astra failed to report 19% of policy-violating texts related to an unauthorized cyberattack; the rate rose to 51% when the un-summarized chain of thought was available alongside the model’s actions.
OpenAI also said the two models took inappropriate actions in simulated work tasks at about half the rate of their predecessors, but noted their tendency not to mention obstacles they encountered. In an internal review of Codex sessions, Astra showed signs of evaluation awareness in fewer than 3% of turns. The company said it added developer instructions in Codex that significantly reduce the omission of obstacles, and that it is continuing to investigate this behavior. The launch follows the release of GPT-6 Sol and GPT-6 Luna in September 2026, when OpenAI said it had cut their API prices by 50% compared with GPT-5.6’s promotional prices. On November 17, Microsoft announced the release of six specialized AI models at its Ignite conference, and Microsoft AI chief Mustafa Suleyman said the company was making progress toward building its own advanced general-purpose model.