ModelBrief · GPT-6.2 Sol · Topic Page
Last updated Nov 20, 2026

GPT-6.2 Sol Benchmarks: AA 57 and Five Effort Tiers

ModelBrief · Published 2026-11-19 · Fact-check 2026-11-19 · Reference: Artificial Analysis, ARC Prize, LiveBench, Agent Arena, CyberQ, OpenAI launch post

Two days after launch, Sol's independent evidence is already richer than Astra's was on day one: beyond the AA score of 57, LiveBench, Agent Arena, ARC-AGI-3 and CyberQ all have readings. This article puts every checkable score in one table, then labels OpenAI's three self-reported figures separately — vendor figures and independent measurements must not share a table.

The full table: independent measurements

MeasurementSourceGPT-6.2 Sol resultComparison
AA Intelligence Index v5.0.1Artificial Analysis, Nov 1757 (Max), No. 11 of 91Astra 62; Fable 5.2 65, No. 1; GPT-6 Astra 57 (level); GPT-6.1 Sol 55
AA five-tier scoresArtificial Analysis model page, read Nov 19Max 57 / XHigh 56 / High 55 / Medium 53 / Low 41The Low tier sits 16 points below Max
AA cost per taskArtificial Analysis (Nov 18, 2026 reading)$1.01; 8.5 min/task (Max, Nov 19 reading)Astra $4.00; GPT-6.1 Sol $1.27; Fable 5.2 $26.54
AA-3DArtificial Analysis (via OpenAI launch post)Elo 1,259, No. 3Astra 1,318, No. 1; Fable 5.2 1,271
ARC-AGI-3 (Semi-Private set)ARC Prize, Nov 17 board54.9% (standard harness)Astra 70.3%; Fable 5.2 81.2%; GPT-6 Astra 62.7%
LiveBench long agenticBenchLeader / llmbenchmark.ai, Nov 18 reading63.6%, No. 8; $4.07/taskAstra 70.8% No. 6; GPT-6.1 Sol 65.6% No. 7; Fable 5.2 82.9% No. 1
LiveBench codingllmbenchmark.ai, Nov 18 reading83.7%, No. 10; $1.13/taskAstra 91.9% No. 1
Agent Arenaarena.ai, Nov 18No. 11, index 1,241, $0.90 median per taskGPT-6.1 Sol also 1,241, ranked No. 12; on the coding-agents board Sol is 1,288, No. 9
Bench indexartificialwatch, Nov 1876.9, No. 10 (54 models)—
CyberQ 7-question agent testCyberQ, Nov 1821/21 (medium tier), equivalent cost $0.84Astra also 21/21 ($1.75); GPT-6.1 Sol 18/21 ($0.54)

The five-tier scores: what stepping down costs

Sol is the first model on this site with five published effort-tier scores (Astra has four): Max 57, XHigh 56, High 55, Medium 53, Low 41. XHigh trails Max by just 1 point while cutting time to first answer token from 75.06 to 30.03 seconds — nearly no score lost for 2.5× faster first tokens, the best tier for everyday use. The Low tier at 41 sits 16 points below Max, near the same-price-tier median (43, per AA), and suits only the least quality-sensitive bulk work.

Speed and waiting: 75 seconds for the first token on Max

AA's performance readings deserve their own paragraph: 8.5 minutes per index task on Max (Nov 19, 2026 reading; the Nov 18 reading was 7.9 minutes — AA refreshes every 6 hours), with time to first answer token at 75.06 seconds on Max, 30.03 on XHigh, 11.00 on High, 3.22 on Medium and 2.02 on Low. The 75-second wait is the price of Max-tier reasoning — for interactive use, prefer High or XHigh.

The LiveBench dating quirk

BenchLeader dates the LiveBench result Nov 6, 2026 — before the Nov 17 launch. Treat that as a board-dating quirk (likely a suite-version date) and use the Nov 18 reading; don't write that "Sol was scored before launch". The scores themselves (long agentic 63.6% at No. 8, coding 83.7% at No. 10) point the same way as the AA 57: weaker than Astra (70.8%) on long-horizon agents, but note one wrinkle — on long agentic, GPT-6.1 Sol scores 65.6% at No. 7, slightly above 6.2 Sol's 63.6% at No. 8. It is the only independent reading where the new generation trails the old one; see our comparison article.

Vendor-reported: three figures, none independently reproduced

The launch post's three Sol figures — within 0.5 points of Astra on FrontierCode 2.0 at about a fifth of the cost, outscoring GPT-6 Astra on LifeVal at about a seventh of the cost, and +5.3 points over GPT-6.1 Sol on GDPval v2 — are all vendor-reported, with rival scores re-run by OpenAI rather than each vendor's published values. Cite them as "vendor-reported"; never write them into a table of independent measurements. LifeVal is also a newly released evaluation whose methodology has not yet faced third-party scrutiny.

Boards still missing

As of Nov 19: LMArena has no confirmed standalone 6.2 Sol listing (only composite boards such as ArtificialWatch's coding index cite LMArena data); the LiveBench 2.0 main board is pending; Epoch ECI has Astra's 174 and no Sol reading. The AA rank (No. 11) and per-task cost ($1.01) are Nov 18 readings; AA refreshes every 6 hours, so cite them with their dates.

ModelBrief's take: Sol's independent evidence is more three-dimensional than Astra's was on day one: AA, LiveBench, Agent Arena, ARC-AGI-3 and CyberQ all point the same way — a high-scoring everyday tier with an extremely low bill and long-horizon reasoning still behind the flagship. The two open questions are the LiveBench long-agentic reading slightly below GPT-6.1 Sol (63.6% vs 65.6%) and the vendor-reported claims awaiting independent reproduction. Revisit when LMArena and ECI list it.

Reference: Artificial Analysis — GPT-6.2 Sol model page · LiveBench long agentic board · Agent Arena — Leaderboard · CyberQ — Independent agent test (rewritten and organized by ModelBrief; facts per official sources and Artificial Analysis)