Inference
EN

Claude Fable 5.2 vs Step 5 Preview

Change models
Claude Fable 5.2AnthropicSmartest
Step 5 PreviewStepFun
IndexAll categories, out of 100.
85.6 ±3.7 best
68.4 ±3.7
GapIndex points below the top.
Top
−17.2
CoverageIndex weight that has results.
90%
90%
SpeedOutput tokens per second.
64/s
105/s best
EloLMArena rating and rank.
—
1443
ReleasedMonth of release.
2026-11
2026-09
Input / 1MUS dollars per 1M input tokens.
$8.00
$1.00 best
Output / 1MUS dollars per 1M output tokens.
$40.00
$2.70 best
Batch output / 1MBatch price per 1M output tokens.
$20.00
—
1K chats2K tokens in, 500 out, each.
$36.00
$3.35 best
1K documents20K tokens in, 1K out, each.
$200.00
$22.70 best
1K agent tasks200K tokens in, 20K out, each.
$2,400.00
$254.00 best
ContextMaximum tokens in one request.
1M
1M
AgenticMulti-step tasks with tools.
82.2 best
68.8
CodingCode writing and repair.
86.2 best
71.4
ReasoningLogic problems and puzzles.
86.5 best
67.4
MultimodalTasks with images and text.
82.0 best
62.2
KnowledgeFacts and expert knowledge.
87.1 best
63.5
MultilingualTasks in many languages.
—
—
InstructionTasks with strict prompt rules.
—
—
MathMath problems.
79.3 best
61.7

Benchmarks

23 shared
Only Claude Fable 5.2: 48Only Step 5 Preview: 39
Claude Fable 5.2
Step 5 Preview
Artificial Analysis Humanity's Last ExamKnowledge
74.3% best1 / 222
46.5%23 / 222
Artificial Analysis Long Context ReasoningReasoning
83.3%13 / 217
88.3% best2 / 217
Critical Physics TasksReasoning
35.4% best1 / 214
20.9%26 / 214
Artificial Analysis Omniscience AccuracyKnowledge
79.1% best1 / 164
36.4%64 / 164
Vibe Code Bench v1.1Coding
93.8% best1 / 123
70.4%48 / 123
Artificial Analysis SciCodeCoding
67.7% best1 / 114
58.9%13 / 114
Legal Research BenchLegal
46.2% best19 / 90
40.4%35 / 90
Tax Agent BenchFinance
32.1% best23 / 82
27.3%39 / 82
Humanity's Last ExamKnowledge
74.2% best1 / 77
46.5%26 / 77
Terminal-Bench 4.0Agentic
65.2% best1 / 61
31.8%23 / 61
DeepSWEAgentic
74.3% best9 / 59
67.7%27 / 59
Agent Arena steerabilityAgentic
33.0 best1 / 54
−0.930 / 54
Agent Arena task outcomeAgentic
22.0 best1 / 54
−1.535 / 54
ProofBench v1.1Math
100.0% best1 / 54
42.0%37 / 54
Terminal-Bench-ScienceScience
55.7% best2 / 44
2.9%33 / 44
HLAB-Vals-1.1Legal
5.0%27 / 35
12.9% best15 / 35
AutomationBenchAgentic
51.4% best11 / 32
44.0%17 / 32
BioMysteryBenchScience
83.7% best1 / 30
67.4%19 / 30
MysteryMechanismScience
51.8% best2 / 30
18.5%27 / 30
Multimodal Knowledge Agent BenchmarkAgentic
56.6% best1 / 19
27.6%19 / 19