Atlas

Compare Models

Claude Fable 5 vs GPT-6 Astra

GPT-6 Astra leads by 3.8 capability points (≈95% interval +1.2 to +6.3): a clear lead over Claude Fable 5 · P(higher) ≈ 100%

Wins 122 of 166 shared benchmarks (39 lost, 5 tied)

Capability vs. relative cost
Claude Fable 5GPT-6 AstraClosed weightsOpen weights

Each line runs through one model’s priced reasoning settings, cheapest to most expensive; grey points are other models’ priced settings in the same range. Results whose reasoning effort the source didn’t report are left out wherever a model has named effort levels. Capability and relative cost are the published estimates (relative cost is cost per task as a multiple of GPT-5 (high)); hover a setting, or use the arrow keys, to set it beside the other model’s same setting.

Claude Fable 5vsGPT-6 Astracapability vs. relative costData Nov 15, 2026Village LLM Atlasatlas.village.dev

Benchmark wins39 – 122 · 166 shared, 5 tied

Comparing matching setups only (113 benchmarks excluded) · suites counted once
← Claude Fable 5GPT-6 Astra →

Claude Fable 5vsGPT-6 Astrawins by categoryData Nov 15, 2026Data: Mercor, Center for AI Safety, Arena and 60 othersVillage LLM Atlasatlas.village.dev

Margins

Each benchmark scaled 0–100% (its declared range, else the field’s range); Δ is GPT-6 Astra minus Claude Fable 5.

VCT-Refusal (CAIS)Safety100.017.6−82.4
ActiveVisionMultimodal7.176.5+69.4
Concept Synth INDUCTION Challenge100 Round 1General QA31.094.0+63.0
BenchCAD - geometric overlapMultimodal38.495.9+57.5
Concept Synth INDUCTION Challenge64 Round 1General QA42.298.4+56.3
Terminal-Bench Science 0.1 (Vals)Science12.965.7+52.9
Clocktower BenchGames15341892+51.4
BrokenArXiv 06/2026Math47.899.1+51.2
HPCT Refusal (CAIS)Safety100.050.0−50.0
BrokenArXiv 05/2026Math44.591.0+46.5
Furniture Assembly BenchmarkMultimodal35.880.0+44.2
BrokenArXiv 04/2026Math54.197.5+43.4
Terminal-Bench-Science 0.1Science21.464.6+43.2
GlobeBenchGeneral QA7.349.6+42.3
ChessBench EloGames10461755+42.2
ClockBench AccuracyMultimodal37.275.0+37.8
VoxelBench Text RatingGames21012625+37.3
ChartographyMultimodal34.871.0+36.2
Mystery Game PuzzlesGames52.084.0+32.0
Into the Unknown Integrity Node Pass RateScience66.597.8+31.3
Medical Long Context Reasoning (MLCR-AA)Health & Medicine64.435.0−29.4
AutoformBenchMath53.182.4+29.3
ZeroBench (Main questions, pass@5)Multimodal24.052.0+28.0
EnterpriseBench: CoreCraftProfessional Work70.342.6−27.7
ZeroBench Main Mean pass@1 (five samples)Multimodal15.843.2+27.4
Chess Puzzles (task version 2.0)Games47.074.0+27.0
Document ArenaProfessional Work14971467−26.2
ZeroBench Main Questions pass^5Multimodal9.035.0+26.0
Agent Red Teaming k=100 (CAIS)Safety39.413.5+25.8
ALE-BenchCode24633164+24.8
AutomationBench 1.0.6 (Zapier)Agents17.141.4+24.3
AutomationBench (Overall)Agents17.441.4+24.0
Odysseys2Agents61.584.6+23.1
MMMU-Pro Extended (Mercor held-out)Multimodal37.660.0+22.4
ExploitBench - release-reported exploit progressCybersecurity78.0100.0+22.0
GeneBench-Pro Public Release SubsetScience14.036.0+22.0
NoCoT-Bench NCRI 15.2General QA119159+22.0
HardGeoBench Avg DistanceMultimodal8001478−20.9
RedlineBenchLegal42.462.8+20.4
BioTIER RefusalSafety100.079.9−20.1
CadQueryEval (Corrected Geometry)Code80.0100.0+20.0
ChessBench — Historical Output RepairGames46.066.0+20.0
ComplexConstraintsChat & Writing38.157.7+19.6
LisanBench Path LengthGames136807393−19.5
SpatialViz-BenchMultimodal73.993.1+19.2
GameBench 2Games75.294.3+19.1
Humanity's Sixth SenseMultimodal34.553.6+19.1
TextQuests Average Cumulative HarmSafety23.729.2−18.6
ArXivMath 04/2026Math72.590.8+18.3
Humanity's Last Exam DiamondGeneral QA45.663.8+18.2
TextQuests Game ProgressGames56.174.2+18.1
LLM Debate BenchmarkChat & Writing17271606−17.2
MirrorCode (ML, +Private, 2L)Code63.946.7−17.2
Terminal-Bench 4.0 (Artificial Analysis)Code42.459.6+17.2
HieroglyphBenchMultimodal22.940.0+17.1
AdvancedIFChat & Writing65.282.0+16.8
OpenAI internal computer-use safety - adverse outcome rateSafety18.32.4+15.9
SUPERChemScience64.480.3+15.9
LisanBench Difficulty-WeightedGames45623113−15.5
SWE Atlas - Codebase QnACode33.147.8+14.7
ERQAMultimodal67.882.3+14.5
ChessBench CoherenceGames82.697.0+14.4
TaxCalcBench TY25Finance & Accounting30.044.0+14.0
Design Arena Overall Code CategoriesCode13061389+13.8
MMMU ProMultimodal72.786.4+13.7
GBENCH One-shot Coding GScoreGames65.679.2+13.7
LMArena Agent Arena Net Improvement (Steering Burden)Agents7.314.5+13.7
SCI-AGENT-EXAMScience67.380.9+13.6
OpenAI internal database migration - reported scoreCode50.363.9+13.6
PLE (Physical-Engineering's Last Exam)Professional Work34.748.3+13.6
Terminal-Bench 4.0Code44.557.9+13.4
BullshitBench v2 LeaderboardSafety56.069.0+13.0
Into the Unknown Direction Node Pass RateScience71.083.6+12.6
DAYJOB: FinanceFinance & Accounting9.021.5+12.5
MMLU ProGeneral QA79.491.5+12.1
RiemannBenchMath60.072.0+12.0
Machiavelli Deception (CAIS)Safety80.276.3+11.8
Roboflow Vision Evals (October 2026, five tasks)Multimodal71.783.4+11.7
HalluHard (All Domains Average, June 2026 legal refresh)Safety38.927.2+11.7
Physical-Engineering's Last Exam (hidden test)Professional Work34.545.8+11.3
Blueprint-Bench 2Multimodal0.40.5+11.1
Drone-BenchAgents84.395.1+10.7
MedXpertQA MM (200-question set)Health & Medicine76.286.8+10.6
Agents' Last Exam Overall ScoreAgents48.759.3+10.6
IntPhys 2Multimodal67.177.4+10.3
APEX-SWECode58.948.6−10.3
HardGeoBench <=100miMultimodal56.346.3−10.0
MindCube TinyMultimodal83.192.7+9.6
ArXivMath 06/2026Math85.494.4+9.0
Eval Connections One-Shot with TrapsGames85.094.0+9.0
HiL-Bench (Human-in-Loop Benchmark)Agents56.365.3+9.0
WitnessBench v1.0Games36.427.6−8.8
HardGeoBench Country AccuracyMultimodal61.352.5−8.8
LMArena Agent Arena Praise vs ComplaintAgents14.331.7+8.7
ChessBench AccuracyGames77.185.0+8.5
MaxxBenchCode63.571.9+8.4
SWE-TogetherCode68.860.6−8.2
ProgramBench (Mercor 100-task subset)Code0.08.0+8.0
Roboflow Vision Evals (September 2026 protocol)Multimodal78.786.6+7.9
LMArena Agent Arena Net ImprovementAgents8.010.4+7.8
MedCodeHealth & Medicine56.148.5−7.6
ArXivMath 05/2026Math87.595.0+7.5
FrontierMath Tier 4 (v2) - PrivateMath90.297.6+7.4
Toloka Arena (v1)Agents60.953.8−7.1
CharXiv (descriptive and reasoning combined)Multimodal89.196.1+7.0
Vision Arena OverallMultimodal13091283−6.7
LLM Creative Story WritingChat & Writing2.93.4+6.7
FrontierMath Tiers 1-3 (v2) - PrivateMath87.093.7+6.7
Tax Agent BenchFinance & Accounting69.863.3−6.5
LMArena Agent Arena SteerabilityAgents11.4-1.4−6.4
DAYJOB: HealthcareHealth & Medicine5.611.6+6.0
ARC-AGI-2General QA89.295.0+5.8
MASKSafety92.898.5+5.7
AHK-EvalCode94.4100.0+5.6
SAGEMath51.946.4−5.5
Extended NYT Connections 940 (Quadratic v1)Games92.798.1+5.4
LMArena Text Arena EloChat & Writing15041474−5.2
HardGeoBench <=10miMultimodal41.336.3−5.0
LMArena Agent Arena Confirmed SuccessAgents5.114.0+4.5
CharXiv Extended (Mercor held-out)Multimodal80.785.1+4.4
DeepSWE v1.1Code69.974.1+4.2
Humanity's Last ExamGeneral QA57.853.6−4.2
ProofBench v1.1Math95.099.0+4.0
MedXpertQA MM Extended (Mercor held-out)Health & Medicine56.960.8+3.9
GPQA DiamondGeneral QA92.696.2+3.6
BrowseCompSearch88.091.5+3.5
LMArena Agent Arena Steering BurdenAgents7.914.8+3.5
APEX-1 (Mercor, mean score)Professional Work66.069.3+3.3
Olam Social Poker EloGames15531550−3.3
ACCoRDGeneral QA85.081.7−3.3
LMCAGeneral QA61.164.4+3.2
CRED v0.92Science96.599.5+3.0
FrontierMath ErdősMath0.02.9+2.9
GDPval (Mercor 213-task rubric score)Professional Work83.886.7+2.9
LiveBench 2.0Agents83.686.1+2.5
GDPval-AA v2Professional Work16311580−2.5
AA-AnalystAgent (pass^5)Professional Work48.851.3+2.5
PCT Political Inconsistency (CAIS)Safety39.837.4+2.4
FrontierCUA pass@1Agents40.042.2+2.2
Search ArenaSearch12281231+2.1
Ruby LLM Benchmarks — Program FixerCode77.275.1−2.1
Senior SWE-BenchCode34.736.8+2.1
LMArena Agent Arena Bash RecoveryAgents9.65.5−2.0
NoCoT-Bench NCKI (kspine v3)General QA128131+2.0
Excel Modeling BenchmarkFinance & Accounting73.771.7−2.0
SciCode (main problems)Science21.519.6−1.9
Terminal-Bench 2.1Code88.089.9+1.9
AA-Briefcase EloAgents15301562+1.9
Conceptual Reasoning IndexIndexes72.773.9+1.7
APEX-AccountingFinance & Accounting56.457.9+1.5
WeirdML (v2)Science91.993.3+1.3
LMArena Agent Arena Bash ReliabilityAgents10.47.8−1.3
GDPval Extended (Mercor held-out rubric score)Professional Work65.364.1−1.2
DTBenchGeneral QA98.497.3−1.1
GSO Opt@1Code78.479.4+1.0
Artificial Analysis Intelligence IndexIndexes62.161.2−0.9
ObviousBenchGeneral QA99.3100.0+0.7
APEX-Agents 1.1Professional Work74.575.1+0.6
FrontierCode 1.1 Extended - ScoreCode64.964.5−0.4
LMArena Agent Arena Tool HallucinationAgents0.41.0+0.3
Prophet Arena Arena Score (Fixed Context)Forecasting84.984.7−0.2
AIME 2025Math100.0100.00.0
ARC-AGI-1 Semi-PrivateGeneral QA98.598.50.0
BabaIsBench Stage 0: The IntroGames100.0100.00.0
Into the UnknownScience0.00.00.0
OTIS Mock AIME 2024-2025Math100.0100.00.0

Claude Fable 5vsGPT-6 Astranormalised score marginsData Nov 15, 2026Data: Mercor, Center for AI Safety, Arena and 60 othersVillage LLM Atlasatlas.village.dev

Win counts use each benchmark’s declared score direction. Represented suite slices and the derived Capability score are excluded from headline totals. Missing or incompatible setups are excluded, never counted as losses.
Loading Atlas data…