Kimi K3 vs Step 5 Preview
Full profiles: Kimi K3 (Moonshot AI) and Step 5 Preview (StepFun): every benchmark, price and rank.
Updated · Data from 13 sources
As of 16 Nov 2026, Kimi K3 is ahead of Step 5 Preview on quality (60.7 vs 56.0 on the BenchLeader Index), but Step 5 Preview is 4.2× cheaper at $1.43 vs $6.00 per 1M tokens blended. Kimi K3 leads in every category we measure both in. Step 5 Preview streams ~1.5× faster.
Key facts (as of 16 Nov 2026)
- Kimi K3 scores 60.7 on the BenchLeader Index and Step 5 Preview scores 56.0 (BenchLeader, 16 Nov 2026).
- Kimi K3 costs $6.00 per 1M tokens blended (list price, 16 Nov 2026).
- Step 5 Preview costs $1.43 per 1M tokens blended (list price, 16 Nov 2026).
- Step 5 Preview streams ~1.5× faster than Kimi K3 (Artificial Analysis, 7-day median, 16 Nov 2026).
Which to pick for…
The pick's numbers come first. Category scores on the BenchLeader scale (50 is the average model, 15 points one standard deviation). High confidence means a lead of half a standard deviation or more; under one point is called either.
| Use case | Pick | Why | Confidence |
|---|---|---|---|
| Coding | Kimi K3 | 62 vs 50 category score; SciCode 58.7% vs 58.9% | High |
| Agents & tool use | Kimi K3 | 54 vs 49 category score | Medium |
| Maths | Kimi K3 | 72 vs 60 category score; ProofBench 87.0% vs 42.0% | High |
| Knowledge | Kimi K3 | 66 vs 54 category score | High |
| Chat quality (human preference) | Kimi K3 | 69 vs 63 on Arena votes | Medium |
| Lowest cost | Step 5 Preview | $1.43 vs $6.00 per 1M blended; agentic coding session $0.071 vs $0.240 | – |
| Fastest first answer | Kimi K3 | 0.8 s (Kimi K3) vs 23.5 s (Step 5 Preview) to the first answer token | – |
| High volume on a budget | Neither flagship | A smaller model is usually the better fit: Cheapest good LLMs, Fastest LLMs | – |
Key differences
- Kimi K3 leads on quality: 60.7 vs 56.0.
- Kimi K3 is stronger in agents & tools, coding, human preference, knowledge, maths.
- Step 5 Preview is 4.2× cheaper ($1.43 vs $6.00 per 1M blended).
- Step 5 Preview streams 1.6× faster (124 vs 76 tokens per second).
At a glance
| Spec | Kimi K3 | Step 5 Preview |
|---|---|---|
| Developer | Moonshot AI | StepFun |
| Released | 16 Jul 2026 | 18 Sep 2026 |
| Context window (tokens) | 1,048,576 | 1,000,000 |
| Compared at | Kimi K3 | Step 5 Preview |
| Open weights | Yes | Yes |
| Price in / out / cached input per 1M | $3.00 / $15.00 / $0.300 | $1.00 / $2.70 / $0.105 |
Category by category
Category scores on the BenchLeader scale: 50 is the average evaluated model, 15 points is one standard deviation. A gap under one point is called level.
| Category | Kimi K3 | Step 5 Preview | Leader |
|---|---|---|---|
| Reasoning | – | 51.7 | not comparable |
| Coding | 61.8 | 49.9 | Kimi K3 |
| Agents & tools | 53.5 | 49.1 | Kimi K3 |
| Maths | 71.5 | 60.1 | Kimi K3 |
| Knowledge | 66.1 | 53.7 | Kimi K3 |
| Human preference | 68.6 | 63.1 | Kimi K3 |
| Multimodal | – | 59.3 | not comparable |
| Long context | – | 69.0 | not comparable |
On one chart
Each domain on the index's own scale, where 50 is an average model. Click a domain to see the benchmarks behind it. Pick up to 5 models from the list.
Not drawn, measured for fewer than two of these models: Reasoning, Instruction following, Long context, Multimodal.
Where sources disagree
Across the 2 benchmarks both have, Kimi K3 leads 1, Step 5 Preview leads 1. Independent sources split: Kimi K3 leads on BenchLeader Index and Arena Text; Step 5 Preview leads on AA Intelligence Index v5.0.1 and Vals Index.
| Overall index | Kimi K3 | Step 5 Preview | Leader |
|---|---|---|---|
| BenchLeader Index | 60.7 | 56.0 | Kimi K3 |
| AA Intelligence Index v5.0.1 | 17.4† | 28.5 | Step 5 Preview |
| Vals Index | 53.0† | 54.8 | Step 5 Preview |
| Arena Text | 1487† | 1449 | Kimi K3 |
Each publisher runs its own tasks, harnesses and effort settings, so their overall verdicts can differ; BenchLeader reconciles them on one scale. † measured at another effort setting of the same model (the publisher did not test the one compared here).
Price, speed and latency
List prices blended at three input tokens per output token. Per-task prices use the token counts on the methodology page. Cost, tokens and time per benchmark task are measured: Artificial Analysis’ averages over its whole Intelligence Index run.
| Measure | Kimi K3 | Step 5 Preview |
|---|---|---|
| Blended price per 1M tokens | $6.00 | $1.43 |
| Input / output per 1M | $3.00 / $15.00 | $1.00 / $2.70 |
| Cached input per 1M | $0.300 | $0.105 |
| Price per task: chat reply | $0.0057 | $0.0012 |
| Price per task: summarize a 30-page report | $0.045 | $0.014 |
| Price per task: code edit | $0.041 | $0.010 |
| Price per task: agentic coding session | $0.240 | $0.071 |
| Price per task: structured extraction | $0.0090 | $0.0025 |
| Agentic coding session, input mostly cached | $0.118 | $0.030 |
| Cost per benchmark task (AA Index) | – | $2.35 |
| Output tokens per task (AA Index) | – | 113k |
| Of which reasoning | – | 80k |
| Time per task (AA Index) | – | 15 min |
| Output speed | 76 tok/s | 124 tok/s |
| Time to first answer | 0.8 s | 23.5 s |
| Time to a complete answer, last measured 13 Nov 2026 | – | 54.4 s |
Each model is compared at its best-ranked reasoning-effort setting. Kimi K3 scores 60.7 and takes 1 s to its first answer token; Kimi K3 (max) scores 60.2 in 55 s, Kimi K3 (low) scores 53.6 in 8 s. Step 5 Preview scores 56.0 and takes 24 s to its first answer token. Higher effort means more thinking before the answer: better scores, longer waits and more tokens billed. How reasoning effort changes scores and cost.
Every benchmark
173 benchmarks, grouped by category; each name links to the board and its leaders. † measured at another effort setting of the same model, shown for reference and not counted as a win.
| Metric | Publisher | Kimi K3 | Step 5 Preview |
|---|---|---|---|
| BenchLeader Index | BenchLeader | 60.7 | 56.0 |
| Agents & tools score | BenchLeader | 53.5 | 49.1 |
| Coding score | BenchLeader | 61.8 | 49.9 |
| Human preference score | BenchLeader | 68.6 | 63.1 |
| Knowledge score | BenchLeader | 66.1 | 53.7 |
| Maths score | BenchLeader | 71.5 | 60.1 |
| Long context score | BenchLeader | – | 69.0 |
| Multimodal score | BenchLeader | – | 59.3 |
| Reasoning score | BenchLeader | – | 51.7 |
| Blended price $/M | BenchLeader | $6.00 | $1.43 |
| Output speed | BenchLeader | 76 tok/s | 124 tok/s |
| Time to first answer | BenchLeader | 0.8 s | 23.5 s |
| Context window | BenchLeader | 1,048,576 tokens | 1,000,000 tokens |
| Reasoning | |||
| GPQA Diamond | Epoch AI Benchmarking Hub | 93.1%† | – |
| SimpleBench | SimpleBench | 60.7%† | – |
| Arena Hard Prompts | Arena (LMArena) | 1516† | 1477 |
| LiveBench 2.0 Reasoning | LiveBench | 90.7%† | 83.8%† |
| LiveBench 1.0 Reasoning | LiveBench | 90.7% | – |
| GPQA Diamond (AA) | Artificial Analysis | 93.5%† | – |
| Humanity's Last Exam (AA) | Artificial Analysis | 46.9%† | 46.5% |
| HLE-Diamond (AA) | Artificial Analysis | 26.9%† | 24.3% |
| GPQA Diamond (Vals) | Vals AI | 92.9% | – |
| ARC-AGI-1 | ARC Prize | 94.5%† | – |
| ARC-AGI-2 | ARC Prize | 60.4%† | – |
| CritPt | Artificial Analysis | 23.4%† | 20.9% |
| CritPt v2 hard subset (AA) | Artificial Analysis | 40.5%† | 32.5% |
| MysteryMechanism | Vals AI | – | 18.5% |
| HLE-Diamond (Scale) | Scale AI SEAL | 22.2% | – |
| Arena Expert | Arena (LMArena) | 1528† | 1489 |
| Chess Puzzles | Epoch AI Benchmarking Hub | 36.0%† | – |
| Mystery Game Puzzles | Epoch AI Benchmarking Hub | 26.0%† | – |
| LMCA | Epoch AI Benchmarking Hub | 52.7%† | – |
| DTBench | Epoch AI Benchmarking Hub | 91.2%† | – |
| ForecastBench | ForecastBench | 64.2%† | – |
| Coding | |||
| SciCode | SciCode | 58.7% | 58.9% |
| WeirdML | WeirdML | 82.6%† | – |
| FrontierCode | Cognition | 44.2% | – |
| Arena Coding | Arena (LMArena) | 1538† | 1505 |
Show all 173 benchmarks
| Metric | Publisher | Kimi K3 | Step 5 Preview |
|---|---|---|---|
| Coding | |||
| Arena WebDev | Arena (LMArena) | 1658† | 1573† |
| LiveBench 2.0 Coding | LiveBench | 53.1%† | 67.6%† |
| LiveBench 1.0 Coding | LiveBench | 81.5% | – |
| SciCode (AA) | Artificial Analysis | 59.5%† | 58.9% |
| LiveCodeBench | Vals AI | 87.2%† | – |
| IOI | Vals AI | 48.9%† | 38.1% |
| SWE-bench (Vals) | Vals AI | 93.4% | – |
| Code Migration | Vals AI | 16.1%† | 39.0% |
| Code Migration v1.1 | Vals AI | 28.6%† | 47.0% |
| ProgramBench | Vals AI | 2.0% | – |
| SWE-bench Multimodal | Vals AI | 44.2%† | 40.8% |
| Vibe Code Bench 1-100 | Vals AI | 18.2%† | – |
| Vibe Code Bench v1.1 | Vals AI | 85.0%† | 70.4% |
| SWE Atlas v1.1 | Scale AI SEAL | 61.0% | – |
| SWE Atlas v1.1: Codebase QnA | Scale AI SEAL | 65.3% | – |
| SWE Atlas v1.1: Test Writing | Scale AI SEAL | 63.3% | – |
| SWE Atlas v1.1: Refactoring | Scale AI SEAL | 54.3% | – |
| SWE-Bench Pro V2 | Scale AI SEAL | 97.7%† | – |
| SWE-Bench Pro V2 HARD | Scale AI SEAL | 88.2%† | – |
| MirrorCode | Epoch AI Benchmarking Hub | 54.4%† | – |
| Surface Evolver Bench | Epoch AI Benchmarking Hub | 95.0% | – |
| DeepSWE v1.1 | Epoch AI Benchmarking Hub | 68.5%† | – |
| ALE-Bench | Epoch AI Benchmarking Hub | 1524.5† | – |
| FrontierSWE | Epoch AI Benchmarking Hub | 25.9%† | – |
| WeirdML v3 | WeirdML | 7.5%† | – |
| Terminal-Bench 4.0 (AA) | Artificial Analysis | 12.6%† | 33.3% |
| Terminal-Bench 2.1 (AA) | Artificial Analysis | 85.0%† | – |
| Startrise Frontend | Startrise | 79.8% | – |
| GameBench 2 Overall | DuelLab GameBench 2 | 48.0 | – |
| GameBench 2 by reasoning setting | DuelLab GameBench 2 | 45.7† | – |
| Design Arena Website | Design Arena | 1336 | – |
| FrontierCode 1.1 (AA) | Artificial Analysis | 43.6%† | 38.4% |
| FrontierCode 2.0 (AA) | Artificial Analysis | 10.9%† | 12.0% |
| Agents & tools | |||
| Terminal-Bench | Terminal-Bench | 23.0%† | – |
| APEX-Agents | Mercor | 47.8% | – |
| Arena Agent | Arena (LMArena) | 0.6† | -3.7 |
| LiveBench Long Agentic | LiveBench | 36.0%† | 19.4%† |
| LiveBench Browser | LiveBench | 75.2%† | 24.8%† |
| LiveBench Agentic Coding | LiveBench | 62.2% | – |
| Terminal-Bench 2.1 (Vals) | Vals AI | 80.9% | – |
| MCP Atlas | Scale AI SEAL | 82.3%† | – |
| HiL-Bench | Scale AI SEAL | 42.7% | – |
| τ³-bench Banking | τ²-bench | 37.1%† | – |
| GDPval-AA v2.1, normalised score | Artificial Analysis | – | 54.3% |
| GDPval-AA v2.1 | Artificial Analysis | 1542† | 1591 |
| τ³-Banking (AA) | Artificial Analysis | 46.0%† | – |
| ITBench SRE (AA) | Artificial Analysis | 47.7%† | 55.6% |
| Analyst Agent (AA) | Artificial Analysis | 38.8%† | 35.0% |
| APEX-Agents (AA) | Artificial Analysis | 41.3%† | 38.0% |
| CyberBench | Vals AI | 75.2%† | – |
| Finance Agent v2 | Vals AI | 53.1%† | 50.7% |
| Finance Agent v2.1 | Vals AI | 60.0%† | 58.9% |
| Harvey's Legal Agent Benchmark v1.1 | Vals AI | 12.9%† | 12.9% |
| Harvey's Legal Agent Benchmark v1 | Vals AI | 12.9%† | 9.2% |
| Legal Research Bench | Vals AI | 46.1%† | 40.4% |
| SREBench | Vals AI | – | 0.4% |
| Tax Agent Bench | Vals AI | 33.4%† | 27.3% |
| Terminal-Bench 4.0 (Vals) | Vals AI | 17.2%† | 31.8% |
| Terminal-Bench Science | Vals AI | 1.4%† | 2.9% |
| Time Horizon Index: KSP | Vals AI | 10.5%† | – |
| Arena Agent Task Completion | Arena (LMArena) | 3.4† | -1.5 |
| PostTrainBench | Epoch AI Benchmarking Hub | 31.5% | – |
| GBAEval | Epoch AI Benchmarking Hub | 48.4% | – |
| Vending-Bench 2 | Andon Labs | 5165.0 | – |
| AutomationBench | Artificial Analysis | 58.3%† | 51.0% |
| GDP.pdf (AA) | Artificial Analysis | 22.0%† | 14.8% |
| Harvey LAB-AA v1.1 | Harvey | 5.3%† | – |
| EnterpriseOps-Gym | Artificial Analysis | 45.3%† | 47.2% |
| AA-Briefcase v1.1 | Artificial Analysis | 1497† | 1424 |
| GDPx | GDPx by SemiAnalysis | 3.7%† | – |
| Design Arena Agentic | Design Arena | 1261 | – |
| Terminal-Bench-Science 0.1 (AA) | Artificial Analysis | 8.1%† | 2.4% |
| CWE-Bench (AA) | Artificial Analysis | 53.3%† | 52.5% |
| DeepSec-Bench (AA) | Artificial Analysis | 12.4%† | 15.8% |
| CyberGym end-to-end (AA) | Artificial Analysis | 58.0%† | 74.8% |
| MKA (AA) | Artificial Analysis | – | 27.6% |
| FACTS Search | 90.3% | – | |
| Maths | |||
| FrontierMath Tiers 1–3 | Epoch AI Benchmarking Hub | 72.2%† | – |
| FrontierMath Tier 4 | Epoch AI Benchmarking Hub | 39.0%† | – |
| OTIS Mock AIME | Epoch AI Benchmarking Hub | 97.2%† | – |
| ProofBench | Vals AI | 87.0% | 42.0% |
| LiveBench Mathematics | LiveBench | 84.4% | – |
| AIME 2026 | MathArena | 96.7%† | – |
| HMMT February 2026 | MathArena | 97.0%† | – |
| MathArena Apex | MathArena | 65.6%† | – |
| Arena Maths | Arena (LMArena) | 1502† | 1470 |
| FrontierMath Erdős | Epoch AI Benchmarking Hub | 0.0%† | – |
| Knowledge | |||
| SimpleQA Verified | Epoch AI Benchmarking Hub | 50.6%† | – |
| LiveBench 2.0 Data Analysis | LiveBench | 78.7%† | 63.5%† |
| LiveBench Research | LiveBench | 79.8%† | 75.5%† |
| LiveBench 1.0 Data Analysis | LiveBench | 78.7% | – |
| AA-Omniscience v1.1 | Artificial Analysis | -1.7† | 16.4 |
| MMLU-Pro | Vals AI | 88.0%† | – |
| LegalBench | Vals AI | 86.2%† | – |
| CorpFin | Vals AI | 71.6% | – |
| TaxEval | Vals AI | 75.7% | – |
| BioMysteryBench | Vals AI | 72.6%† | 67.4% |
| Excel Modeling Benchmark | Vals AI | 66.7%† | 57.1% |
| Excel Modeling Benchmark v1.1 | Vals AI | 71.3%† | 69.5% |
| MedCode | Vals AI | 49.4%† | – |
| MedScribe | Vals AI | 88.0%† | – |
| Public Benefits Bench v1.1 | Vals AI | 68.2%† | 58.1% |
| Access to Justice Benchmark | Vals AI | 68.2%† | – |
| Financial Coaching Bench | Vals AI | 74.5%† | – |
| Immigration Bench | Vals AI | 68.3%† | – |
| GDP.pdf | Epoch AI Benchmarking Hub | 19.0%† | – |
| AA-Omniscience v1.1: accuracy | Artificial Analysis | 42.4%† | 36.4% |
| AA-Omniscience v1.1: non-hallucination | Artificial Analysis | 23.4%† | 68.7% |
| HealthBench Professional (AA) | OpenAI | 53.2%† | – |
| Instruction following | |||
| LiveBench Language | LiveBench | 85.5% | – |
| LiveBench Instruction Following | LiveBench | 71.4% | – |
| Arena Instruction Following | Arena (LMArena) | 1489† | 1450 |
| Human preference | |||
| Arena Text | Arena (LMArena) | 1487† | 1449 |
| EQ-Bench 4 | EQ-Bench | 1319 | – |
| EQ-Bench Creative Writing v3 | EQ-Bench | 2083 | – |
| Arena Creative Writing | Arena (LMArena) | 1459† | 1413 |
| Arena Multi-turn | Arena (LMArena) | 1502† | 1445 |
| Arena Longer Queries | Arena (LMArena) | 1500† | 1461 |
| Multimodal | |||
| Arena Vision | Arena (LMArena) | – | 1263 |
| LiveBench Design | LiveBench | 72.5%† | 70.8%† |
| MMMU-Pro | Artificial Analysis | 80.5%† | 76.4% |
| MMMU-Pro (Vals) | Vals AI | 88.2% | – |
| MortgageTax | Vals AI | 66.3% | – |
| SAGE | Vals AI | 86.8%† | – |
| Humanity's Sixth Sense | Scale AI SEAL | 25.5%† | – |
| Blueprint-Bench 2 | Epoch AI Benchmarking Hub | 29.5% | – |
| Furniture Assembly | Epoch AI Benchmarking Hub | 34.2%† | – |
| Code2Video-AA | Artificial Analysis | 1041† | 821 |
| AA-3D | Artificial Analysis | 984† | 859 |
| Long context | |||
| AA-LCR v1.1 | Artificial Analysis | 88.7%† | 88.3% |
| MLCR | Artificial Analysis | 21.7%† | 18.3% |
| Composite | |||
| Epoch Capabilities Index | Epoch AI Benchmarking Hub | 158.6 | – |
| LiveBench 2.0 | LiveBench | 65.0%† | 54.9%† |
| LiveBench 1.0 | LiveBench | 79.2% | – |
| AA Intelligence Index v5.0.1 | Artificial Analysis | 17.4† | 28.5 |
| AA Intelligence Index v4.3.2 | Artificial Analysis | 43.6† | 43.7 |
| AA Progress Index v1 | Artificial Analysis | 319.9† | 318.2 |
| Vals Index | Vals AI | 53.0† | 54.8 |
| Vals Finance Index | Vals AI | 58.0%† | 55.4% |
| Vals RSI Index | Vals AI | 20.9%† | – |
| Vals Science Index | Vals AI | – | 29.6% |
| Vals Multimodal Index | Vals AI | 73.4% | – |
| Vals Router Index | Vals AI | 53.0% | – |
| Vals Legal Index | Vals AI | 71.7%† | 70.2% |
| AA Cyber Index | Artificial Analysis | 41.3† | 47.7 |
| Safety & honesty | |||
| FORTRESS | Scale AI SEAL | 27.3% | – |
| DistressBench | Scale AI SEAL | 73.5% | – |
| Metadata | |||
| AA Openness Index | Artificial Analysis | 38.9† | 33.3 |
Data as of 2026-11-16. Best configuration of each model; every score links to its source on the model pages.
DuelLab GameBench 2, release gb2-20261111-7b598bb13de4 (https://benchmarks.duellab.org/releases/gb2-20261111-7b598bb13de4/), CC BY 4.0 (https://creativecommons.org/licenses/by/4.0/). Scores mapped to BenchLeader model names; not endorsed by DuelLab.
What changed
- Kimi K3: BenchLeader Index 59.1 → 60.7.
- Step 5 Preview: BenchLeader Index 54.2 → 56.0.
- Step 5 Preview: BenchLeader Index 54.7 → 54.2.
- Kimi K3: 1 score revision.
- Step 5 Preview: 1 score revision.
- Kimi K3: 4 score revisions.
- Step 5 Preview: 7 score revisions.
- Kimi K3: 1 score revision.
- Kimi K3: 1 new result (FACTS Search).
- Kimi K3: 1 score revision.
Related comparisons
- Kimi K3 vs Claude Opus 5.5
- Step 5 Preview vs Claude Opus 5.5
- Step 5 Preview vs GPT-6 Astra
- Step 5 Preview vs GPT-6.1 Sol
- Step 5 Preview vs Gemini 4 Argon
- Kimi K3 vs GPT-6.1 Sol
- Kimi K3 vs GPT-5.6 Sol
- Kimi K3 vs Gemini 4 Argon
Rankings: Best LLM overall · Best LLM for agents · Best cheap LLMs
Kimi K3 vs Step 5 Preview: questions
- Is Kimi K3 better than Step 5 Preview?
- Kimi K3 leads on quality: 60.7 vs 56.0. The BenchLeader Index combines every independent quality benchmark; Kimi K3 is ahead overall as of 16 Nov 2026, but check the category scores for your use.
- Kimi K3 vs Step 5 Preview: which is cheaper, and will that change?
- Step 5 Preview is cheaper today: $1.43 against $6.00 per million tokens, blended at three input tokens per output token. No price change has been announced for either.
- Is Kimi K3 or Step 5 Preview better for coding?
- Kimi K3 scores higher in coding (62 vs 50 on the category index, where 50 is average). On SciCode, the top shared coding benchmark, it is 58.7% for Kimi K3 against 58.9% for Step 5 Preview.
- Which is faster, Kimi K3 or Step 5 Preview?
- Step 5 Preview streams ~1.5× faster (124 vs 76 output tokens per second, Artificial Analysis 7-day median).
- Is Kimi K3 or Step 5 Preview better for agents and tool use?
- Kimi K3 scores higher in agentic tasks (54 vs 49 on the category index, where 50 is average).
More questions (4)
- Which is better for long documents?
- Both take about 1M tokens: Kimi K3 1,048,576, Step 5 Preview 1,000,000.
- Which has lower latency?
- Kimi K3 answers first: 0.8 s against 23.5 s to the first answer token, as Artificial Analysis measures it. Reasoning effort drives most of this: a higher setting thinks longer before it answers.
- Do independent benchmarks agree on Kimi K3 vs Step 5 Preview?
- Across the 2 benchmarks both have, Kimi K3 leads 1, Step 5 Preview leads 1. Independent sources split: Kimi K3 leads on BenchLeader Index and Arena Text; Step 5 Preview leads on AA Intelligence Index v5.0.1 and Vals Index.
- When was Step 5 Preview released?
- Step 5 Preview was released on 18 Sep 2026.