baseline
Claude 2.1 → Claude 3
31.9 → 50.4
GPQA moved from low-thirties to roughly half correct by the first Opus generation.
Claude’s highest reported GPQA score rose 59.4 points in 27 months
Across nine major Claude releases with recoverable GPQA Diamond results, the reported high moved from 31.9% for Claude 2.1 to 91.3% for Opus 4.6. The biggest visible step came with Claude 3.7 Sonnet, though changing evaluation settings mean this is a history of published results—not a controlled rerun.
91.3%
latest selected high
+59.4
percentage-point gain
2.9×
score multiple
9
release points traced
Reported GPQA Diamond score by releasehighest recovered result, %
20406080100
Claude 2.1
Nov ’23
Nov ’23
3 Sonnet
Mar ’24
Mar ’24
3 Opus
Mar ’24
Mar ’24
3.5 Sonnet
Jun ’24
Jun ’24
3.7 Sonnet
Feb ’25
Feb ’25
Opus 4
May ’25
May ’25
Sonnet 4
May ’25
May ’25
Opus 4.5
Nov ’25
Nov ’25
Opus 4.6
Feb ’26
Feb ’26
What changed
1
The Claude 3 family established the first clear staircase. Sonnet’s 40.4% was followed by Opus at 50.4% under the reported GPQA Diamond results. langbase.com
2
Claude 3.7 Sonnet produced the largest milestone jump. Its highest recovered result was 84.8%, 19.8 points above the selected 3.5 Sonnet result. vellum.ai
3
Not every later model raised the line. Opus 4’s recovered 79.6% sat below 3.7 Sonnet’s high, while Sonnet 4 reached 83.8% with a deeper-thinking/tools setting. anthropic.com
4
Opus regained the frontier. Opus 4.5 reached 87.3%, followed by 91.3% for Opus 4.6 in the selected published result. claude5.ai
Three eras
reasoning jump
Claude 3.5 → 3.7
65.0 → 84.8
Reasoning modes drove the sharpest reported increase, but also reduced protocol comparability.
frontier consolidation
Opus 4.5 → 4.6
87.3 → 91.3
The selected Opus results pushed published performance above nine in ten questions.
Release detail
| Date | Model | Score | Reported setting | Source |
|---|---|---|---|---|
| 2023-11-21 | Claude 2.1 | 31.9% | Not recovered | aiflashreport.com |
| 2024-03-01 | Claude 3 Sonnet | 40.4% | 0-shot CoT | langbase.com |
| 2024-03-04 | Claude 3 Opus | 50.4% | 0-shot CoT in family table | promptlayer.com |
| 2024-06-01 | Claude 3.5 Sonnet | 65.0% | Standard mode | getpassionfruit.com |
| 2025-02-24 | Claude 3.7 Sonnet | 84.8% | Highest of two reported settings | vellum.ai |
| 2025-05-22 | Claude Opus 4 | 79.6% | Highest recovered headline value | anthropic.com |
| 2025-05-22 | Claude Sonnet 4 | 83.8% | Deep thinking + tools | anthropic.com |
| 2025-11-24 | Claude Opus 4.5 | 87.3% | Not recovered | claude5.ai |
| 2026-02-05 | Claude Opus 4.6 | 91.3% | Standard/no extended thinking in recovered source | agentwiki.org |
Method: web search run 25 Aug 2026 across Anthropic pages, benchmark aggregators and model-analysis sites; 1,052 unique search results were scanned and 30 normalized Claude-name candidates were assembled. Values are the highest recovered published GPQA Diamond result for each selected release, not a fresh uniform evaluation. Dates and protocols occasionally differ across sources, so compare the long-run direction more confidently than small adjacent gaps.