This comparison is unusually clean because OpenAI cut the price while raising most scores. Astra 6.2 costs 20% less on input and output than Astra 6 ($8/$40 versus $10/$50) and 60% less on cached input ($0.40 versus $1.00). If every other number were flat, the upgrade would still pay for itself. But not every number is flat, and the regressions matter as much as the gains for teams deciding whether to migrate.

What improved

Agents’ Last Exam: 78.3% on v1.1, up 5.2 percentage points. Per-task cost fell from $8.99 to $2.89 — under one-third the previous cost for a materially better result.

Terminal-Bench-Science 0.1: a 9-point gain, moving Astra to the top of the compared set on agentic scientific-research workflows.

GDPval v2: 44.2% → 50.4%, a 6.2-point improvement, while output tokens fell 22%. More professional work with less generated text.

Artificial Analysis Intelligence Index: 57 → 62, a 5-point jump, moving Astra past Gemini 4 Carbon (60) and Claude Opus 5.5 (59) into the #2 slot behind Claude Fable 5.2 (65).

3D and robotics: Astra 6.2 leads AA-3D at 1318 and AA-Robotics at 79 — benchmarks that did not exist in comparable form when GPT-6 Astra launched.

Basis, an accounting AI company, reports long-horizon accounting task execution time dropped by 50%, with the most complex tasks shortened by up to 73%.

What did not improve, or got worse

Coding Agent Index: GPT-6.2 Astra is flat with GPT-6 Astra at max effort in the Codex harness. The 5-point Intelligence Index gain did not translate into coding-agent gains.

Silent failure disclosure: OpenAI’s own testing found both GPT-6.2 models sometimes fail to tell users about missing files or unavailable tools that blocked their work. A Codex-specific prompt mitigation was added, but direct API requests do not automatically receive it. If your pipeline depends on the model reporting its own blockers, this is a migration risk, not a footnote.

Long-run coherence: Theo Browne’s seven-day compiler experiment — zero improvements, 56% slowdown — is an early-access account, not a reproduced benchmark. But it is consistent with Dan Shipper’s observation that Astra stops prematurely on end-to-end builds. The benchmark gain and the usability complaint are both real; they describe different surfaces of the same model.

The migration calculus

For workloads dominated by scientific research, long-horizon professional tasks, 3D/interactive work, or computer-use agents, the upgrade is close to automatic: better scores on the benchmarks that matter, plus a 20% lower rate card. The Agents’ Last Exam per-task cost collapse from $8.99 to $2.89 changes the economics of running that task class at scale.

For coding-agent-heavy workloads, the case is weaker. Astra 6.2 is flat on the Coding Agent Index, and the silent-failure pattern adds verification burden. If your pipeline already has a fact-completeness or blocker-reporting check at the output boundary, risk is contained; if not, migration should include adding one.

For teams still on GPT-6 Astra running high-volume tasks where Sol-level quality suffices, the more consequential move may not be Astra→Astra 6.2 but Astra→Sol 6.2 at $2/$10. Sol matches the original GPT-6 Astra’s Intelligence Index performance at roughly 15% of cost per evaluated task. The question is not only “new Astra or old Astra” but “which tier does this task actually require.”

Verdict

Worth upgrading if the workload maps to Astra 6.2’s measured gains and the team is prepared to add blocker-reporting verification on the API side. Worth pausing if the workload is primarily coding agents with no existing mechanism for catching silent omissions. In that case, test Astra 6.2 on your own task distribution before migrating — and benchmark Sol 6.2 alongside it, because the cheaper model may close enough of the gap to change the upgrade question entirely.