Inference
EN
StepFun

Step 5 Preview

Step 5 Preview is a reasoning model from StepFun in the Step 5 family. 37 benchmarks count toward its index, in 6 categories.

  • Open weights
  • Reasoning
  • Input: text · image · video
  • Released 2026-09
RankedEnough results for an index.
IndexAll categories, out of 100.68.4±3.7Median 51.0
CoverageIndex weight that has results.90%
SpeedOutput tokens per second.105/sMedian 67/s
First tokenSeconds to the first output token.3.5 sMedian 1.3 s
Input / 1MUS dollars per 1M input tokens.$1.00Median $0.60
Output / 1MUS dollars per 1M output tokens.$2.70Median $2.55
ContextMaximum tokens in one request.1MMedian 262K
EloLMArena rating and rank.1443±6 #70Median 1422

Details

From OpenRouter
API IDThe model ID on OpenRouter.
stepfun/step-5-preview
Max outputMaximum output tokens per request.
64K
Knowledge cutoffLast date of training data.
—
ToolsTool calls on OpenRouter.
Yes
JSON outputOutput follows a given JSON schema.
Yes
Cache read / 1MUS dollars per 1M cached tokens.
$0.10
WeightsLink to the published weights.
Hugging Face (opens in a new tab) ↗
RetiresDate OpenRouter removes the model.
—

Where to run

APIs that serve this model.
4 providers
ProviderTypeMaker, cloud, host or router.InputOutputContext
Deep InfraDeprecatedHost$1.00$2.701.05M
Novita AIHost$1.00$2.701M
OpenRouterRouter$1.00$2.701M
StepFunMaker$1.00$2.701.02M

US dollars per 1M tokens, list price. From models.dev (opens in a new tab) (MIT).

Run locally

604.34B parameters
16-bit
1270.1 GB
8-bit
675.2 GB
4-bit
385.7 GB
Weights
Hugging Face (opens in a new tab) ↗

Memory need with a short context. The 4-bit copy needs more than 128 GB. Compare local models

Capabilities

Index per category, out of 100.
Out of 100
AgenticMulti-step tasks with tools.22 counted68.828 / 218
CodingCode writing and repair.6 counted71.417 / 209
ReasoningLogic problems and puzzles.3 counted67.436 / 227
MultimodalTasks with images and text.2 counted62.236 / 84
KnowledgeFacts and expert knowledge.3 counted63.550 / 268
MultilingualTasks in many languages.None counted——
InstructionTasks with strict prompt rules.None counted——
MathMath problems.1 counted61.7—

Results

37 counted
BenchmarkThe benchmark name.ResultThe score from the source.PlacePlace among results on this benchmark.
Artificial Analysis Long Context ReasoningReasoning88.3%2 / 217
Scientific Code BenchmarkCoding58.9%2 / 35
MCP AtlasAgentic85.6%4 / 52
Artificial Analysis ITBench-AAAgentic55.6%2 / 24
Artificial Analysis Humanity's Last ExamKnowledge46.5%23 / 222
Artificial Analysis SciCodeCoding58.9%13 / 114
GPQA Diamond · BenchLMKnowledge93.5%10 / 84
Critical Physics TasksReasoning20.9%26 / 214
APEX-Agents-AAAgentic38.0%5 / 30
GDPval-AA normalizedAgentic54.3%24 / 133
BrowseCompAgentic88.7%12 / 63
Terminal-Bench 2.1 · BenchLMAgentic85.0%20 / 86
ProgramBench: Can Language Models Rebuild Programs From Scratch?Coding80.5%5 / 19
Artificial Analysis CyberGym-E2E-AAAgenticAbove fitFar above this model's index.74.8%5 / 19
CyberGymAgentic84.7%8 / 29
JobBenchAgentic59.0%11 / 36
Humanity's Last ExamKnowledge46.5%26 / 77
Toolathlon-VerifiedAgentic74.1%8 / 23
Artificial Analysis Cyber IndexAgentic47.7%7 / 19
Terminal-Bench 4.0Agentic31.8% ±2.623 / 61
Artificial Analysis MMMU-ProMultimodal76.4%44 / 113
Vibe Code Bench v1.1Coding70.4% ±4.348 / 123
Artificial Analysis Omniscience AccuracyKnowledge36.4%64 / 164
Code MigrationCoding39.0% ±4.531 / 77
LiveBench Coding 2.0Coding85.6%22 / 52
DeepSWEAgentic67.7%27 / 59
LiveBench DesignAgentic76.9%24 / 52
APEX-AgentsAgentic37.8%6 / 13
LiveBench ResearchAgentic72.8%27 / 52
AutomationBenchAgentic44.0%17 / 32
LiveBench ReasoningReasoning83.8%44 / 80
Agent Arena steerabilityAgentic−0.930 / 54
Massive Multi-discipline Multimodal Understanding ProMultimodal76.0%33 / 53
Agent Arena command recoveryAgentic0.232 / 50
Agent Arena task outcomeAgentic−1.535 / 54
Agents' Last ExamAgentic29.5%14 / 21
Artificial Analysis DeepsecBench-AAAgentic15.8%14 / 21
Artificial Analysis CWE-Bench-AAAgentic52.5%13 / 19
ProofBench v1.1Math42.0% ±5.037 / 54
OfficeQA ProMultimodal60.3%11 / 16
LiveBench Long AgenticAgentic27.5%36 / 51
LiveBench Browser UseAgentic24.0%41 / 49
IOICoding38.1% ±6.738 / 45
LiveBench Data AnalysisReasoning63.5%69 / 80
Hermes Index: agentic capability and cost per task inside Hermes AgentAgentic32.8%14 / 16
Multimodal Knowledge Agent BenchmarkAgentic27.6%19 / 19
Legal Research BenchLegal40.4% ±3.435 / 90
HLAB-Vals-1.1Legal12.9% ±2.815 / 35
Tax Agent BenchFinance27.3% ±3.139 / 82
Finance Agent (v2)Finance50.7% ±0.342 / 79
EMBFinance57.1% ±3.241 / 75
BioMysteryBenchScience67.4% ±0.419 / 30
Public Benefits Bench v1.1Public benefits58.1% ±1.337 / 50
Terminal-Bench-ScienceScience2.9% ±2.033 / 44
SRE BenchOther0.4% ±0.431 / 36
MysteryMechanismScience18.5% ±2.627 / 30
Artificial Analysis Harvey LAB-AA v1.0Agentic93.4%—
Data Research and Analysis with Complex OperationsAgentic83.3%—
SWE-MarathonCoding72.7%—
Hermes Bench: 150 agent tasks graded on the files and state the agent leaves behindAgentic69.2%—
Artificial Analysis EnterpriseOps-GymAgentic47.2%—
MLS-Bench LiteCoding40.5%—

Sources

BenchLMCC BY-NC 4.0OpenRouterTerms restrict reuseLiveBenchNo licence statedVals AINo licence statedLMArenaCC BY 4.0models.devMIT

History

5 releases
Index68.0 → 68.4
23 Oct 202612 Nov 2026
Output / 1M$2.70 → $2.70
23 Oct 202612 Nov 2026

Questions

How much does Step 5 Preview cost?
$1.00 per 1M input tokens, $2.70 per 1M output tokens.
How good is Step 5 Preview?
Its index is 68.4 out of 100. That is place 29 of 283 ranked models.
What is Step 5 Preview best at?
Coding: 71.4 out of 100, place 17 of 209.
How fast is Step 5 Preview?
105 output tokens per second, 3.5 s to the first token.
What is the context window of Step 5 Preview?
1M tokens.
Is Step 5 Preview open source?
Yes. Its weights are published.
When was Step 5 Preview released?
September 2026.

Badge

Step 5 Preview on Inference360: place 29 of 283, index 68.4
HTML
<a href="https://inference360.ai/models/step-5-preview"><img src="https://inference360.ai/badges/step-5-preview.svg" alt="Step 5 Preview on Inference360: place 29 of 283, index 68.4" height="20"></a>
Markdown
[![Step 5 Preview on Inference360: place 29 of 283, index 68.4](https://inference360.ai/badges/step-5-preview.svg)](https://inference360.ai/models/step-5-preview)

Compare with

Gemini 3.7 Flash68.5DeepSeek V4.1 Flash 101768.5GPT-5.6 Terra68.9Gemini 3.8 Flash67.8Grok 4.667.8

Lower price

Ling 3.1 Flash67.2 · $0.135/1MDeepSeek V4.1 Flash 101768.5 · $0.40/1MClaude Haiku 5.566.3 · $0.50/1M

More from StepFun

All models
Step 3.7 Flash55.3Step 3.5 FlashUnrankedStep-AudioUnrankedStep-Audio-Chat 130BUnranked