The benchmark board

Every score from our Model Drops from the last 30 days on one board, since Oct 6, 2026: who leads, how fast the best scores move, and which new models jumped furthest past the last generation.

60benchmarks

21models scored

307scores logged

10Model Drops

Best model overall

Handy AI's weigh-in: every leaderboard on this page folded into one ranking.

Best score over time

Each dot is a model at its release. The line is the record holder at each point. Hollow dots are models without a drop of their own, placed where we first saw them cited.

Closed weights only

AA-Omniscience hallucination rate

Best score fell 28.3 points, from 51.3% (GPT-6 Astra) to 23% (Muse Wonder Full).

Closed weights only

AutomationBench-AA

Best score rose 8.3 points, from 69.5% (Claude Opus 5.5) to 77.8% (Gemini 4 Carbon).

Closed weights only

FrontierCode 2.0

Best score rose 7.4 points, from 34.9% (Claude Haiku 5.5) to 42.3% (GPT-6.1 Terra).

Closed weights only

CritPt v2 Hard

Best score rose 25.5 points, from 43.5% (Claude Haiku 5.5) to 69% (Gemini 4 Carbon).

Closed weights only

Terminal-Bench Science (standard harness)

Best score rose 44.8 points, from 20% (Claude Haiku 5.5) to 64.8% (Gemini 4 Carbon).

Biggest generational jumps

A new model against the model it replaced, on the same chart in its launch post.

Claude Haiku 4.5Claude Haiku 5.5

OSWorld 2 (offline subset): 15.7% to 72.4%

+56.7

GPT-5.6 TerraGPT-6.1 Terra

AA-Omniscience hallucination rate: 88% to 38%

+50

Leaderboards

The latest reported score for every model on each benchmark. Tap a card's share button for an image of it.

Closed weights only

Terminal-Bench 4.0

Multi-hour work in a real terminal.

Claude Sonnet 5.5max 70.6%

Claude Opus 5.5max 66.4%

GPT-6 Astramax 58.2%

Fable 5.1max 57.9%

Muse Wonder Fullmax 57.1%

GPT-6.1 Solmax 55.9%

10 more

Muse Wondermax 52%

Claude Opus 5 51.8%

GPT-6.1 Terramax 50.3%

GPT-5.6 Sol 39.9%

Claude Haiku 5.5max 39.2%

GPT-5.6 Terramax 21.5%

GPT-6 Luna 16.4%

Fern 0.1 6.1%

Muse Glimmer-30Bmax 1.6%

Claude Haiku 4.5 0%

16 models · 11 drops · GPT-6.1 Terra, Kimi K3.1 Pro, Muse Wonder, MiniMax M3.1 Flash, Claude Haiku 5.5, Mirendil, "Introducing Fern", Nanbeige4.3-33B model card

Closed weights only

DeepSWE v1.1

Fixing issues in real repositories.

Muse Wonder Fullmax 79%

Muse Wondermax 78.8%

Claude Opus 5.5max 74.2%

GPT-6 Astramax 74.1%

Claude Opus 5 74%

GPT-5.6 Sol 73%

4 more

Particle-1max 72.6%

Fable 5.1max 67.4%

Fern 0.1 44.2%

Muse Glimmer-30Bmax 4.4%

10 models · 9 drops · Muse Wonder, Kimi K3.1 Pro, MiniMax M3.1 Flash, Particle, "Introducing Particle-1", Mirendil, "Introducing Fern", Nanbeige4.3-33B model card

Closed weights only

GPQA Diamond

5 models compared.

GPT-6 Astramax 96.1%

Claude Opus 5.5max 93.9%

Fable 5.1max 93.7%

Particle-1 90.4%

Muse Glimmer-30Bmax 83.5%

5 models · 7 drops · Kimi K3.1 Pro, Particle, "Introducing Particle-1", Nanbeige4.3-33B model card

Closed weights only

AutomationBench

Real business automations (Zapier).

Claude Opus 5 50.3%

GPT-5.6 Sol 45.8%

Muse Glimmer-30Bmax 15.2%

3 models · 7 drops · MiniMax M3.1 Flash, Nanbeige4.3-33B model card

Closed weights only

Terminal-Bench 2.1

2 models compared.

Particle-1max 79.9%

Muse Glimmer-30Bmax 61.8%

2 models · 7 drops · Particle, "Introducing Particle-1", Nanbeige4.3-33B model card

Closed weights only

Artificial Analysis Intelligence Index

14 models compared.

Claude Opus 5.5 58

Claude Sonnet 5.5max 56

Muse Wonder Fullmax 55.8

GPT-6 Astra 53

Fable 5.1 53

Gemini 4 Argonhigh 53

8 more

Muse Wondermax 52.4

GPT-6.1 Solmax 52

Claude Opus 5 50.8

Muse Spark 1.3 48.1

GPT-6.1 Terramax 48

Claude Haiku 5.5max 43

GPT-5.6 Terramax 42

GPT-6 Lunamax 38

14 models · 6 drops · Kimi K3.1 Pro, GPT-6.1 Terra, Muse Wonder, DeepSeek V4.1 Flash 1017

Closed weights only

Humanity's Last Exam (no tools)

Expert-level questions, no tools.

Claude Opus 5.5max 59.8%

Fable 5.1max 57.6%

Claude Sonnet 5.5 56.9%

Claude Opus 5 56.3%

GPT-6 Astramax 53.6%

Claude Haiku 5.5max 45.9%

2 more

GPT-5.6 Sol 44.5%

Claude Haiku 4.5 10.2%

8 models · 5 drops · Kimi K3.1 Pro, Claude Haiku 5.5, DeepSeek V4.1 Flash 1017

Closed weights only

ProgramBench

7 models compared.

Claude Opus 5.5max 65%

Muse Wonder Fullmax 59%

Fable 5.1max 53.5%

Muse Wondermax 50%

GPT-6 Astramax 50%

Claude Opus 5 37%

1 more

GPT-5.6 Sol 23%

7 models · 5 drops · Kimi K3.1 Pro, Muse Wonder, DeepSeek V4.1 Flash 1017

Closed weights only

GDPval-AA

Professional work products, graded head to head (Elo).

Claude Opus 5.5max 1866 Elo

Muse Wonder Fullmax 1857 Elo

Claude Sonnet 5.5 1838 Elo

Muse Wondermax 1770 Elo

Fable 5.1max 1754 Elo

Claude Haiku 5.5max 1620 Elo

3 more

GPT-6 Astramax 1573 Elo

GPT-6 Luna 1437 Elo

Claude Haiku 4.5 735 Elo

9 models · 4 drops · Kimi K3.1 Pro, Muse Wonder, Claude Haiku 5.5

Closed weights only

Agents' Last Exam

4 models compared.

Claude Opus 5.5max 42.6%

GPT-6 Astramax 32.8%

Claude Opus 5 32.2%

GPT-5.6 Sol 30.6%

4 models · 4 drops · Kimi K3.1 Pro, MiniMax M3.1 Flash

Closed weights only

Sentiment Bench

How people received the model in its first week on X, Hacker News, in the news and among trusted reviewers (Handy AI).

Gemini 4 Carbon 85.8

Claude Haiku 5.5 71.4

GPT-6.1 Terra 69.2

Muse Wonder 37

4 models · 4 drops · Every model's sources, quotes and method

Closed weights only

SWE-bench Verified

2 models compared.

GPT-6 Astra 86%

Fern 0.1 73.4%

2 models · 4 drops · Mirendil, "Introducing Fern"

Closed weights only

OSWorld 2

Using a computer: apps, files and the web.

Claude Sonnet 5.5max 83.9%

GPT-6 Astramax 80.9%

GPT-6.1 Solmax 78%

GPT-6.1 Terramax 74.2%

Claude Opus 5 70.6%

GPT-5.6 Sol 65.7%

2 more

GPT-5.6 Terramax 57.1%

GPT-6 Lunamax 55.5%

8 models · 3 drops · GPT-6.1 Terra, MiniMax M3.1 Flash

Closed weights only

AA-LCR

6 models compared.

Muse Wonder Fullmax 83.7%

Muse Glimmer-30Bmax 83.3%

GPT-6.1 Terramax 82.7%

Claude Haiku 5.5max 82.7%

Muse Wondermax 80.7%

Gemini 4 Carbonhigh 79.3%

6 models · 3 drops · Artificial Analysis (2026-11-04), Nanbeige4.3-33B model card

Closed weights only

BabyVision (with tools)

6 models compared.

Claude Opus 5.5max 97.7%

GPT-6 Astramax 96.6%

Fable 5.1max 95.1%

Claude Opus 5 94.1%

Particle-1 89.2%

GPT-5.6 Sol 88.9%

6 models · 3 drops · Kimi K3.1 Pro, DeepSeek V4.1 Flash 1017, Particle, "Introducing Particle-1"

Closed weights only

Harvey Legal Agent Benchmark

Legal work end to end.

Fable 5.1max 19.1%

Muse Wonder Fullmax 13.1%

Qwen3.8 Max 10.4%

Muse Wondermax 10%

GPT-6 Astramax 8.6%

Claude Opus 5.5max 4.2%

6 models · 3 drops · Kimi K3.1 Pro, Muse Wonder, Mistral Large 4

Closed weights only

Terminal-Bench Science

Scientific computing tasks in a terminal.

GPT-6 Astramax 68.1%

Claude Opus 5.5max 63.3%

GPT-6.1 Terramax 58.3%

GPT-6.1 Solmax 57%

Fable 5.1max 40%

5 models · 3 drops · GPT-6.1 Terra, Kimi K3.1 Pro

Closed weights only

ZeroBench (with tools)

5 models compared.

Claude Opus 5.5max 61%

Fable 5.1max 59%

GPT-6 Astramax 58%

GPT-5.6 Sol 53%

Claude Opus 5 52%

5 models · 3 drops · Kimi K3.1 Pro, DeepSeek V4.1 Flash 1017

Closed weights only

NL2Repo

3 models compared.

Claude Opus 5 75.3%

GPT-5.6 Sol 56.8%

Muse Glimmer-30Bmax 27.4%

3 models · 3 drops · MiniMax M3.1 Flash, Nanbeige4.3-33B model card

Closed weights only

OSWorld 2 (offline subset)

Using a computer, on the 82 OSWorld 2.1 tasks that need no internet.

Claude Opus 5.5max 87.2%

Claude Sonnet 5.5 83.9%

Gemini 4 Carbonmax 81.5%

GPT-6 Astramax 80.9%

Fable 5.1max 76.8%

Gemini 4 Argonmax 74.7%

3 more

Claude Haiku 5.5max 72.4%

GPT-6 Luna 48.9%

Claude Haiku 4.5 15.7%

9 models · 2 drops · Gemini 4 Carbon, Claude Haiku 5.5

Closed weights only

Terminal-Bench 4.0 (standard harness)

Terminal-Bench run by a third party on one shared harness.

Gemini 4 Carbonhigh 65.2%

Claude Opus 5.5max 59.6%

GPT-6 Astramax 59.1%

Gemini 4 Argonmax 57.1%

Muse Wonder Fullmax 57.1%

GPT-6.1 Terramax 55.6%

3 more

Fable 5.1max 52%

Muse Wondermax 52%

Claude Haiku 5.5max 32.8%

9 models · 2 drops · Artificial Analysis (2026-11-04), Gemini 4 Carbon

Closed weights only

AA-Omniscience hallucination rate

Lower is better.

Muse Wonder Fullmax 23%

Muse Wondermax 26.5%

GPT-6.1 Terra 38%

Claude Sonnet 5.5 47%

GPT-6 Astra 51.3%

Claude Opus 5.5 58.6%

2 more

Fable 5.1 72.6%

GPT-5.6 Terra 88%

8 models · 2 drops · Muse Wonder, GPT-6.1 Terra

Closed weights only

AutomationBench-AA

Business automations, run by Artificial Analysis.

Gemini 4 Carbonhigh 77.8%

Muse Wonder Fullmax 71.3%

Claude Opus 5.5max 69.5%

GPT-6 Astramax 68.5%

Muse Wondermax 68.1%

GPT-6.1 Terramax 60.2%

2 more

Fable 5.1max 59.4%

Claude Haiku 5.5max 35.4%

8 models · 2 drops · Artificial Analysis (2026-11-04), Kimi K3.1 Pro

Closed weights only

Chartography (with tools)

5 models compared.

GPT-6 Astramax 91%

Claude Opus 5.5max 89%

Fable 5.1max 88.4%

Claude Opus 5 84%

GPT-5.6 Sol 79.9%

5 models · 2 drops · Kimi K3.1 Pro, DeepSeek V4.1 Flash 1017

Closed weights only

SWE-Marathon

5 models compared.

Muse Wonder Fullxhigh 57.5%

Muse Wondermax 53.1%

Claude Opus 5.5max 46.3%

Fable 5.1max 45.6%

GPT-6 Astramax 42.5%

5 models · 2 drops · Muse Wonder, Kimi K3.1 Pro

Closed weights only

Humanity's Last Exam (with tools)

Expert-level questions, with search and code.

Claude Opus 5.5max 67.7%

Fable 5.1max 65.6%

GPT-6 Astramax 57.2%

3 models · 2 drops · Kimi K3.1 Pro

Closed weights only

BrowseComp

2 models compared.

GPT-6 Astramax 91.5%

Claude Opus 5.5max 90.8%

2 models · 2 drops · Kimi K3.1 Pro

Closed weights only

SWE-Atlas-QnA

2 models compared.

Fern 0.1 53.2%

Muse Glimmer-30Bmax 17.7%

2 models · 2 drops · Mirendil, "Introducing Fern", Nanbeige4.3-33B model card

From a single launch

Closed weights only

Artificial Analysis Intelligence Index v5.0

11 models compared.

Gemini 4 Carbonhigh 60

Claude Opus 5.5 58

GPT-6 Astra 57

GPT-6.1 Sol 55

Claude Sonnet 5.5 54

GPT-6.1 Terramax 51.3

5 more

Muse Wonder Fullmax 50.1

Gemini 4 Argon 50

Fable 5.1 49

Muse Wondermax 43.6

Claude Haiku 5.5max 36.5

11 models · 1 drop · Artificial Analysis (2026-11-04), Gemini 4 Carbon

Closed weights only

AA-Briefcase v1.1

8 models compared.

Claude Opus 5.5max 1803 Elo

Muse Wonder Fullmax 1729 Elo

Fable 5.1max 1672 Elo

Muse Wondermax 1659 Elo

Gemini 4 Carbonhigh 1610 Elo

Claude Haiku 5.5max 1574 Elo

2 more

GPT-6 Astramax 1569 Elo

GPT-6.1 Terramax 1401 Elo

8 models · 1 drop · Kimi K3.1 Pro, Artificial Analysis (2026-11-04)

Closed weights only

Vals Index v2.2

8 models compared.

Gemini 4 Carbon 70.5%

Claude Opus 5.5max 70.2%

Gemini 4 Argonmax 69.2%

Muse Wonder Full 66.8%

GPT-6 Astramax 65%

Muse Wonder 64.9%

2 more

GPT-6.1 Terra 60.5%

Claude Haiku 5.5 58.7%

8 models · 1 drop · Vals Index v2.2 (2026-11-04), Gemini 4 Carbon

Closed weights only

Agents' Last Exam v1

6 models compared.

Claude Opus 5.5max 63.2%

GPT-6.1 Terramax 60%

GPT-6 Astramax 59.3%

GPT-6.1 Solmax 58.6%

GPT-6 Lunamax 50.9%

GPT-5.6 Terramax 50.7%

6 models · 1 drop · GPT-6.1 Terra

Closed weights only

Artificial Analysis Coding Agent Index v2.0

5 models compared.

Claude Opus 5.5max 62

Claude Sonnet 5.5max 59

GPT-6.1 Solmax 54

GPT-6.1 Terramax 48

GPT-6 Lunamax 31

5 models · 1 drop · GPT-6.1 Terra

Closed weights only

Factual error rate

Lower is better.

GPT-6 Astramax 4.2%

GPT-6.1 Terramax 4.3%

GPT-6.1 Solmax 4.4%

GPT-6 Lunamax 18.4%

GPT-5.6 Terramax 22.3%

5 models · 1 drop · GPT-6.1 Terra

Closed weights only

FrontierCode 2.0

Real coding tasks in open-source repos (Cognition), run by Artificial Analysis.

GPT-6.1 Terramax 42.3%

Gemini 4 Carbonhigh 40.6%

Muse Wonder Fullmax 37.8%

Claude Haiku 5.5max 34.9%

Muse Wondermax 27.9%

5 models · 1 drop · Artificial Analysis (2026-11-04)

Closed weights only

FrontierSWE v2

Frontier software engineering tasks.

GPT-6 Astramax 65.5%

Claude Opus 5.5max 62.3%

Gemini 4 Carbonmax 61.2%

Fable 5.1max 56.3%

Gemini 4 Argonmax 55%

5 models · 1 drop · Gemini 4 Carbon

Closed weights only

MemoryBench

5 models compared.

Muse Wondermax 85.6%

Muse Wonder Fullmax 85.5%

Claude Opus 5.5max 81.5%

GPT-6 Astramax 81.4%

Fable 5.1max 79%

5 models · 1 drop · Muse Wonder

Closed weights only

MMMU-Pro Vision

5 models compared.

GPT-6 Astramax 86.7%

GPT-6.1 Solmax 86.3%

GPT-6.1 Terramax 85.5%

GPT-5.6 Terramax 81.7%

GPT-6 Lunamax 79.9%

5 models · 1 drop · GPT-6.1 Terra

Closed weights only

Personal Superintelligence Index

5 models compared.

Muse Wonder Fullmax 59.1

Claude Opus 5.5max 55.2

GPT-6 Astramax 54.5

Muse Wondermax 54.3

Fable 5.1max 49.9

5 models · 1 drop · Muse Wonder

Closed weights only

SWE-sweep

5 models compared.

Gemini 4 Carbonmax 9.6%

GPT-6 Astramax 9.5%

Gemini 4 Argonmax 4.3%

Claude Opus 5.5max 3.6%

Fable 5.1max 2.9%

5 models · 1 drop · Gemini 4 Carbon

Closed weights only

Agents' Last Exam v1.1

4 models compared.

Gemini 4 Carbonmax 77.4%

GPT-6 Astramax 73.1%

Claude Opus 5.5max 71.6%

Gemini 4 Argonmax 69%

4 models · 1 drop · Gemini 4 Carbon

Closed weights only

Chartography

4 models compared.

Claude Sonnet 5.5 61.6%

Claude Haiku 5.5max 46.4%

GPT-6 Luna 29.1%

Claude Haiku 4.5 6.4%

4 models · 1 drop · Claude Haiku 5.5

Closed weights only

APEX-Agents

3 models compared.

Claude Opus 5.5max 73.5%

Fable 5.1max 68.6%

GPT-6 Astramax 64.7%

3 models · 1 drop · Kimi K3.1 Pro

Closed weights only

Artificial Analysis Coding Agent Index

3 models compared.

Claude Opus 5.5 62

Gemini 4 Carbon 59

Gemini 4 Argon 53

3 models · 1 drop · Gemini 4 Carbon

Closed weights only

BenchCAD (with tools)

3 models compared.

Claude Opus 5.5max 96.2%

GPT-6 Astramax 95.9%

Fable 5.1max 92.6%

3 models · 1 drop · Kimi K3.1 Pro

Closed weights only

ExploitBench

3 models compared.

GPT-6 Astramax 100%

Claude Opus 5.5max 91%

Fable 5.1max 83%

3 models · 1 drop · Kimi K3.1 Pro

Closed weights only

Frontier Finance

3 models compared.

Fable 5.1max 60.8%

Claude Opus 5.5max 56.7%

GPT-6 Astramax 45%

3 models · 1 drop · Kimi K3.1 Pro

Closed weights only

FrontierCode v1.1

Hard, long coding tasks.

Claude Sonnet 5.5xhigh 52.1%

Claude Haiku 5.5max 46.4%

GPT-6 Luna 42.4%

3 models · 1 drop · Claude Haiku 5.5

Closed weights only

GDP.pdf

Work products built from PDFs (Surge AI).

GPT-6 Astramax 34.2%

Claude Opus 5.5max 30.6%

Fable 5.1max 27.6%

3 models · 1 drop · Kimi K3.1 Pro

Closed weights only

MMMU-Pro

3 models compared.

Claude Opus 5.5max 87.7%

GPT-6 Astramax 86.9%

Fable 5.1max 86.3%

3 models · 1 drop · Kimi K3.1 Pro

Closed weights only

OfficeQA Pro v2

3 models compared.

Claude Opus 5.5max 62.2%

Fable 5.1max 54.5%

GPT-6 Astramax 54.4%

3 models · 1 drop · Kimi K3.1 Pro

Closed weights only

OSWorld

3 models compared.

Claude Opus 5.5max 81.8%

Fable 5.1max 80.7%

GPT-6 Astramax 75.8%

3 models · 1 drop · Kimi K3.1 Pro

Closed weights only

AA-Omniscience

2 models compared.

Gemini 4 Carbon 54

Claude Opus 5.5 48

2 models · 1 drop · Gemini 4 Carbon

Closed weights only

Agents' Last Exam (CLI)

2 models compared.

Claude Opus 5.5max 40.7%

GPT-6 Astramax 38.8%

2 models · 1 drop · Kimi K3.1 Pro

Closed weights only

Apex Bench v1.1

2 models compared.

Claude Opus 5 65.8%

GPT-5.6 Sol 54.3%

2 models · 1 drop · DeepSeek V4.1 Flash 1017

Closed weights only

Artificial Analysis Cyber Index

2 models compared.

GPT-6 Astra 33

Claude Opus 5.5 29

2 models · 1 drop · Mistral Large 4

Closed weights only

BankerToolBench

2 models compared.

Claude Opus 5 76.2

GPT-5.6 Sol 74.6

2 models · 1 drop · MiniMax M3.1 Flash

Closed weights only

SWE-bench Pro v2 Hard

2 models compared.

Claude Opus 5.5 100%

Fern 0.1 76.5%

2 models · 1 drop · Mirendil, "Introducing Fern"

Closed weights only

CritPt v2 Hard

Unpublished research-level physics problems, hard subset, run by Artificial Analysis.

Gemini 4 Carbonhigh 69%

GPT-6.1 Terramax 68%

Muse Wonder Fullmax 56.5%

Muse Wondermax 50.5%

Claude Haiku 5.5max 43.5%

5 models · 0 drops · Artificial Analysis (2026-11-04)

Closed weights only

Terminal-Bench Science (standard harness)

Terminal-Bench Science run by a third party on one shared harness.

Gemini 4 Carbonhigh 64.8%

GPT-6.1 Terramax 57.6%

Muse Wonder Fullmax 37.1%

Muse Wondermax 25.7%

Claude Haiku 5.5max 20%

5 models · 0 drops · Artificial Analysis (2026-11-04)

Release pace

Every Model Drop by month, marked with its maker.

6 Oct 2026

How this board works

Scores come from the benchmark section of each handyai.news Model Drop, mostly the makers' launch tables. When drops disagree, the newest drop wins, since later launches re-run rivals on the current version of a benchmark; the model's own launch number is kept as launch_score. Settings (effort level, harness) vary between makers, so compare within a board with care. Models without their own drop are dated by the first drop that cites them.

The data is open: /api/benchmarks (JSON), /benchmarks.md (Markdown), one benchmark at /api/benchmarks/{benchmark}, the best model overall at /api/benchmarks/weigh-in (or /weigh-in/{type} for a topic), a PNG of any card by adding .png, and the list_benchmarks, get_benchmark and get_weigh_in tools on our MCP server.