The scores that moved

Early reviewers reported that GPT-6.2 was uneven and sometimes stopped too soon on builds. On Artificial Analysis’s Intelligence Index at Max, Astra rose from 56.85 to 62.14 and Sol rose from 54.61 to 57.26. ChatGPT Work starts supported consumer plans on Sol Light, which OpenAI maps to Low. Artificial Analysis scored Sol 40.55 at Low.

On Artificial Analysis’s Coding Agent Index in Codex at Max, Astra moved from 53.1 to 53.5 and Sol fell from 52.2 to 48.6. Sol’s codebase question score went from 59.9% to 46.2%. Artificial Analysis attributed the drop to more answers leaving out key facts.

In OpenAI’s deliberate blocker test, GPT-6.2 Astra omitted a known coding blocker in 22.28% of blocked trials, compared with 11.09% for GPT-6 Astra. OpenAI says new instructions in Codex reduced omissions in its evaluation. Direct API requests do not include those instructions.

Benchmaxxed tests

We compared the new models with their predecessors at High effort on a code explanation, a handoff with a missing file and a small webpage edit. Across 3 paired trials per task, all 18 answers passed blind scoring and every blocked handoff flagged the missing file. On the explanation and handoff tasks, the new models finished in 1.4–2.0 seconds, compared with 5.9–8.4 seconds for GPT-6.1 Sol and 6.6–11.4 seconds for GPT-6 Astra. The new models also cost less per answer on those tasks.

For a marble-run webpage, both models advanced the marble along a fixed track on the first prompt. A second prompt required a free marble and real collisions. In 1 attempt each at High, Astra took 153.40 seconds and cost $0.429, and its marble survived a sideways nudge. Sol took 81.43 seconds and cost $0.131, and its marble cleared a jump and landed in the cup.

Where to switch

Switch for everyday questions and small edits, and use Astra for hard problems where you can check the result. Wait on Sol as a coding agent. On long agent runs, ask what the model could not do and check the answer against the log.