NeuBench
Back to timeline
Audited Model Intelligence

Frontier Model Benchmarks.
Living Capability Radar.

Side-by-side empirical performance across 266 frontier reasoning, coding, and open-weights models. Every score cites official technical reports or LMSYS evaluations.

266Models Audited
10Benchmarks
100%Verified Sources
General Knowledge

Artificial Analysis Intelligence Index

% Score
Leader:Claude Opus 5.5(57.6%)+1.6% over Claude Sonnet 5.5
1Claude Opus 5.5
57.6%Indep.
2Claude Sonnet 5.5
56.0%Indep.
3Claude Fable 5.1
53.4%Indep.
4GPT-6 Astra
52.7%Indep.
Coding

Artificial Analysis Coding Index

% Score
Leader:Claude Fable 5.1(81.6%)+4.7% over GPT-6 Astra
1Claude Fable 5.1
81.6%Indep.
2GPT-6 Astra
76.9%Indep.
3Claude Sonnet 5.5
Not reported
—
4Claude Opus 5.5
Not reported
—
Agentic

Artificial Analysis Agentic Index

% Score
Leader:Claude Fable 5.1(57.9%)+6.9% over GPT-6 Astra
1Claude Fable 5.1
57.9%Indep.
2GPT-6 Astra
51.0%Indep.
3Claude Sonnet 5.5
Not reported
—
4Claude Opus 5.5
Not reported
—
Coding

SWE-bench Pro

% Score
Leader:Claude Opus 5.5(89.9%)+8.6% over Claude Sonnet 5.5
1Claude Opus 5.5
89.9%Official
2Claude Sonnet 5.5
81.3%Official
3Claude Fable 5.1
81.2%Official
4GPT-6 Astra
Not reported
—
Agentic

Terminal-Bench 4.0

% Score
Leader:Claude Sonnet 5.5(70.6%)+4.2% over Claude Opus 5.5
1Claude Sonnet 5.5
70.6%Official
2Claude Opus 5.5
66.4%Official
3GPT-6 Astra
57.9%Official
4Claude Fable 5.1
55.8%Official
Reasoning

ARC-AGI-2

% Score
Leader:GPT-6 Astra(95.0%)+3.3% over Claude Opus 5.5
1GPT-6 Astra
95.0%Indep.
2Claude Opus 5.5
91.7%Indep.
3Claude Fable 5.1
90.0%Official
4Claude Sonnet 5.5
Not reported
—
Reasoning

GPQA Diamond

% Score
Leader:GPT-6 Astra(96.0%)Sole evaluated model in selection
1GPT-6 Astra
96.0%Official
2Claude Sonnet 5.5
Not reported
—
3Claude Opus 5.5
Not reported
—
4Claude Fable 5.1
Not reported
—
Reasoning

Humanity's Last Exam

% Score
Leader:Claude Opus 5.5(64.4%)+3.5% over Claude Fable 5.1
1Claude Opus 5.5
64.4%Official
2Claude Fable 5.1
60.9%Official
3Claude Sonnet 5.5
56.9%Official
4GPT-6 Astra
Not reported
—
Mathematics

FrontierMath v2 (Tiers 1-3)

% Score
Leader:GPT-6 Astra(93.7%)+2.5% over Claude Opus 5.5
1GPT-6 Astra
93.7%Indep.
2Claude Opus 5.5
91.2%Indep.
3Claude Fable 5.1
90.2%Indep.
4Claude Sonnet 5.5
88.8%Indep.
Human Preference

Design Arena ELO

ELO
Leader:Claude Opus 5.5(1477)+3 ELO over Claude Sonnet 5.5
1Claude Opus 5.5
1477Indep.
2Claude Sonnet 5.5
1474Indep.
3Claude Fable 5.1
1393Indep.
4GPT-6 Astra
Not reported
—

Responsible Benchmark Interpretation

Evaluation Methodology & Contamination Safeguards
01

Workload Specialization

Leading on coding evaluations (SWE-bench Pro, Terminal-Bench) does not imply superior creative writing or conversational nuance. Choose models tailored to your specific application workload.

02

Latency vs Compute Costs

Frontier reasoning models consume extended test-time compute. Compare the Artificial Analysis Intelligence Index against output price to find the best ROI for high-volume or interactive workloads.

03

Contamination & Satiation

Legacy benchmarks (MATH Level 5, GSM8K, HumanEval, MMLU-Pro) have saturated or stopped being reported. This catalogue tracks the boards labs still publish: SWE-bench Pro, Terminal-Bench 4.0, FrontierMath v2, ARC-AGI-2, GPQA Diamond and HLE.

All 266 frontier model benchmarks independently audited against published technical reports and LMSYS leaderboards.
Verified JSON Telemetry
Back to NeuBench Timeline