GPT-6.2 Astra's cost per task has been disclosed as $2.89, compared with $60.05 for Claude Fable 5.2. Our judgment: competition among foundation models is shifting from one-off capability demonstrations to unit task costs and continuous delivery performance.

According to the source article, GPT-6.2 Astra and Sol have appeared in the web versions of Work and Codex. Official examples cover automotive websites, 3D models, presentation updates, financial analysis, and engineering assembly. In the internal GDPval v2 evaluation, Astra's completion rate rose from 44.2% to 50.4%, while average output tokens fell by roughly 22%; Sol's completion rate increased from 41.5% to 46.9%.

What this is

This is not simply another chat model being added. It is designed to handle longer office workflows: reading old templates and new financial reports, producing editable Excel files, completing valuation analysis, and placing status updates, owners, and review items in the same tracking sheet. Web pages and 3D presentations remain visually compelling, but the more important development is that the models are beginning to reduce search and rework.

Lower costs matter just as much. In professional-task evaluations, Astra costs $2.89 per task, roughly one-twentieth of Claude Fable 5.2. However, per-task cost is not the same as the total cost of a real project; failed reruns, human intervention, and review time must still be included.

Industry view

We believe this release sends a clear signal: when enterprises purchase models, they are moving from evaluating leaderboard scores to asking how much a completed job costs, how long it takes, and how many checks it requires. The Basis report found that average completion time for long-horizon accounting tasks was cut in half, while ClickUp observed roughly 30% fewer review rounds, showing that adoption priorities have shifted toward process efficiency.

Risks are already visible in early feedback. Some developers say the new model still produces unclear explanations or stops partway through tasks, leaving people to clean up the results. Every official demonstration is a selected case, and the 3D engineering animation only shows that the visualization was completed; it does not prove that the work has been validated in a real engineering environment. The more autonomous a model becomes, the more enterprises need to retain approval checkpoints, process logs, and result reviews. The ability to generate something cannot be equated directly with the ability to deliver it.

Impact on regular people

For enterprise IT: When evaluating models, teams should track per-call cost, completion rate, human rework, and review rounds together. Only if GPT-6.2 can consistently lower all of these metrics is it likely to enter high-frequency office workflows.

For individual professionals: Tasks such as organizing financial materials, updating presentations, analyzing files, and creating visualization proposals will become faster. But delivery standards will not disappear automatically, and users must still verify critical data.

For the consumer market: Cheaper, more capable models will drive more software products to add automated execution features. At the same time, mid-task stops and result deviations could become new sources of customer complaints.