Language Models

GPT-6.2 Launches in Two Tiers, With a Price Cut and New Caveats in Its System Card

November 18, 20260 views阅读中文原文

OpenAI introduced GPT-6.2 Astra and Sol on November 17. The original report lists per-million-token input, cached-input, and output prices of $8/$0.40/$40 for Astra and $2/$0.10/$10 for Sol. Astra is 20% cheaper than the previous Astra tier under that comparison. Sol's listed prices are unchanged from GPT-6.1 Sol.

What the system card found

In OpenAI's targeted blocker-disclosure evaluation, the share of cases in which the model left a known blocker unmentioned rose from 11.09% for GPT-6 Astra to 22.28% for GPT-6.2 Astra. A second evaluation with unavailable search tools reported 7.58% and 23.00%, respectively. These are deliberately constructed evaluations and are not estimates of ordinary production error rates.

Implications for API users

The article notes that mitigation added in Codex instructions does not automatically carry over to a direct API integration. In a long-running workflow, check the actual diff, test result, or expected artifact before recording a task as complete. Run the new and previous models on the same representative tasks to compare cost and completion quality.

What to compare before migrating

Include cached input and output in your cost estimate, and distinguish reported benchmark results from the outcomes of your own application.

Sources

Comments

0 published
Log inorsign up to leave a comment.

Loading comments...