"gpt 6.2 sol vs 6.1 sol" appeared in Google autocomplete on the first day of the GPT-6.2 launch — the most searched-for Sol comparison. The answer is clean: the price didn't move at all, the AA index is 2 points higher, per-task cost is down about 20%, and CyberQ went from 18/21 to 21/21. But one independent reading went backwards, and this article doesn't hide it.
| Dimension | GPT-6.2 Sol | GPT-6.1 Sol |
|---|---|---|
| AA Intelligence Index v5.0.1 | 57 (Max) | 55 |
| Price (input / output) | $2 / $10 | $2 / $10 (identical) |
| Cost per index task | $1.01 | $1.27 |
| Time per task | 8.5 min (Max, Nov 19 reading) | 23.1 min |
| Output speed | — | 49 tok/s |
| LiveBench long agentic | 63.6% (No. 8) | 65.6% (No. 7) |
| Agent Arena index | 1,241 (No. 11) | 1,241 (No. 12) |
| GDPval v2 (vendor-reported) | Up 5.3 points over the previous generation | — |
| CyberQ 7-question test | 21/21 ($0.84) | 18/21 ($0.54) |
| Knowledge cutoff | June 1, 2026 | — |
| Ultrafast | Ultrafast first on Pro 200 | — |
| Context | 1.05M tokens | — (not checked in this round) |
The most striking row is time per task: 8.5 minutes against 23.1 — the new Sol runs the AA index task nearly 3× faster, which is the main reason the per-task cost fell ($1.01 vs $1.27), not just the unchanged rate card. The AA index is only 2 points higher, but tasks finish faster and cheaper.
CyberQ's home-made 7-question agent bench (Codex CLI, medium tier, three runs per question) has the new Sol at 21/21 and the old one at 18/21, with the whole gap on question 03 — finding the root cause in a failed vLLM startup log, which needs two DeepGEMM environment variables at once; the previous generation missed the MoE one all three times, while the new Sol named both all three times. It is the most concrete generational improvement: not how much the score rose, but that the previous generation's blind spot is fixed. Equivalent cost is $0.84 for the new Sol against $0.54 for the old — about 56% more expensive, for 3 more points; whether that trade is worth it depends on whether your work includes that kind of two-variable root-cause task.
Honestly stated: on LiveBench long agentic, the new Sol scores 63.6% (No. 8), slightly below the old generation's 65.6% (No. 7). It is the only independent reading where the new generation trails the old one. Two possible reasons: first, BenchLeader dates the LiveBench result Nov 6, 2026 (before launch day — the dating is questionable); second, long-horizon agentic work was never Sol's strength (Astra 70.8%, Fable 82.9%). Either way, know before upgrading: long multi-turn work is not what this generation improves.
The launch post claims Sol improves 5.3 points over GPT-6.1 Sol on GDPval v2 — vendor-reported, with no independent reproduction. Against the independent readings (AA +2 points, CyberQ +3 questions, per-task cost −20%), the direction is credible, but don't cite the number itself as an independent measurement.
Upgrade now: API users already on GPT-6.1 Sol — same price, higher scores, much faster runs; there is no cost to upgrading, only gains, and subscription usage rose from 15–160 to 17–180 messages per 5 hours. Wait: if your main work is LiveBench-style long-horizon multi-turn tasks — the new Sol scores slightly below the old one there, so wait for the next round of readings. Upgrade as a bonus: Pro 200 subscribers — Sol is the first model to bring Ultrafast to that plan, so upgrading adds a speed tier.
Reference: Artificial Analysis — GPT-6.2 Sol model page · CyberQ — Independent agent test · OpenAI — Launch post "Introducing GPT-6.2 Astra and Sol" (rewritten and organized by ModelBrief; facts per official sources and Artificial Analysis)