DeepSeek V4.1 Flash 1017 vs Step 5 Preview
Full profiles: DeepSeek V4.1 Flash 1017 (DeepSeek) and Step 5 Preview (StepFun): every benchmark, price and rank.
Updated · Data from 10 sources
As of 6 Nov 2026, DeepSeek V4.1 Flash and Step 5 Preview are neck and neck on quality (59.2 vs 54.9 on the BenchLeader Index, inside the ±5.7 margin), and DeepSeek V4.1 Flash is 2.7× cheaper at $0.525 vs $1.43 per 1M tokens blended. Choose DeepSeek V4.1 Flash for coding (64 vs 50), agents & tools (63 vs 51), and reasoning (57 vs 53); choose Step 5 Preview for fast first answers (25.9 s vs 52.2 s) and knowledge (55 vs 49) and long context (69 vs 66).
Key facts (as of 6 Nov 2026)
- DeepSeek V4.1 Flash scores 59.2 on the BenchLeader Index and Step 5 Preview scores 54.9 (BenchLeader, 6 Nov 2026).
- DeepSeek V4.1 Flash costs $0.525 per 1M tokens blended (list price, 6 Nov 2026).
- Step 5 Preview costs $1.43 per 1M tokens blended (list price, 6 Nov 2026).
- Of the 40 benchmarks both models have, DeepSeek V4.1 Flash leads 31 (BenchLeader, 6 Nov 2026).
DeepSeek V4.1 Flash 1017 benchmarks 
Step 5 Preview benchmarks 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 | DeepSeek V4.1 Flash 1017 | 64 vs 50 category score; FrontierCode 2.0 (AA) 28.5% vs 12.0% | High |
| Agents & tool use | DeepSeek V4.1 Flash 1017 | 63 vs 51 category score; Arena Agent 2.4 vs -2.9 | High |
| Maths | DeepSeek V4.1 Flash 1017 | 62 vs 60 category score | Medium |
| Reasoning | DeepSeek V4.1 Flash 1017 | 57 vs 53 category score; HLE-Diamond (AA) 25.3% vs 24.3% | Medium |
| Knowledge | Step 5 Preview | 55 vs 49 category score; AA-Omniscience v1.1 16.4 vs 0.9 | Medium |
| Long documents | Step 5 Preview | 69 vs 66 category score; AA-LCR v1.1 88.3% vs 82.7% | Medium |
| Images and other media | DeepSeek V4.1 Flash 1017 | 63 vs 59 category score; MMMU-Pro 80.6% vs 76.4% | Medium |
| Chat quality (human preference) | – | Only Step 5 Preview has an Arena rating (63 category score) | Not comparable |
| Lowest cost | DeepSeek V4.1 Flash 1017 | $0.525 vs $1.43 per 1M blended; agentic coding session $0.023 vs $0.071 | – |
| Fastest first answer | Step 5 Preview | 25.9 s (Step 5 Preview) vs 52.2 s (DeepSeek V4.1 Flash 1017 (max)) 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
- DeepSeek V4.1 Flash 1017 (max) and Step 5 Preview are neck and neck on quality (59.2 vs 54.9, within the ±5.7 margin).
- DeepSeek V4.1 Flash 1017 (max) is stronger in agents & tools, coding, maths, multimodal, reasoning.
- Step 5 Preview is stronger in knowledge, long context, composite, human preference.
- DeepSeek V4.1 Flash 1017 (max) is 2.7× cheaper ($0.525 vs $1.43 per 1M blended).
At a glance
| Spec | DeepSeek V4.1 Flash 1017 | Step 5 Preview |
|---|---|---|
| Developer | DeepSeek | StepFun |
| Released | 17 Oct 2026 | 18 Sep 2026 |
| Context window (tokens) | 1,048,576 | 1,000,000 |
| Compared at | DeepSeek V4.1 Flash 1017 (max) | Step 5 Preview |
| Open weights | Yes | Yes |
| Price in / out / cached input per 1M | $0.300 / $1.20 / $0.0060 | $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 | DeepSeek V4.1 Flash 1017 (max) | Step 5 Preview | Leader |
|---|---|---|---|
| Reasoning | 56.7 | 52.7 | DeepSeek V4.1 Flash 1017 |
| Coding | 64.0 | 50.2 | DeepSeek V4.1 Flash 1017 |
| Agents & tools | 63.0 | 50.6 | DeepSeek V4.1 Flash 1017 |
| Maths | 61.6 | 59.7 | DeepSeek V4.1 Flash 1017 |
| Knowledge | 48.6 | 55.2 | Step 5 Preview |
| Human preference | – | 63.2 | not comparable |
| Multimodal | 63.1 | 59.4 | DeepSeek V4.1 Flash 1017 |
| Long context | 66.2 | 69.2 | Step 5 Preview |
| Composite | – | 40.8 | 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: Instruction following, Human preference.
Where sources disagree
Across the 40 benchmarks both have, DeepSeek V4.1 Flash 1017 leads 31, Step 5 Preview leads 7 and 2 are tied. Every overall index that rates both puts DeepSeek V4.1 Flash 1017 ahead.
| Overall index | DeepSeek V4.1 Flash 1017 (max) | Step 5 Preview | Leader |
|---|---|---|---|
| BenchLeader Index | 59.2 | 54.9 | neck and neck |
| AA Intelligence Index v5.0 | 32.6 | 28.5 | DeepSeek V4.1 Flash 1017 |
| Vals Index | 61.3 | 54.8 | DeepSeek V4.1 Flash 1017 |
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 | DeepSeek V4.1 Flash 1017 (max) | Step 5 Preview |
|---|---|---|
| Blended price per 1M tokens | $0.525 | $1.43 |
| Input / output per 1M | $0.300 / $1.20 | $1.00 / $2.70 |
| Cached input per 1M | $0.0060 | $0.105 |
| Price per task: chat reply | $0.0005 | $0.0012 |
| Price per task: summarize a 30-page report | $0.0043 | $0.014 |
| Price per task: code edit | $0.0036 | $0.010 |
| Price per task: agentic coding session | $0.023 | $0.071 |
| Price per task: structured extraction | $0.0008 | $0.0025 |
| Agentic coding session, input mostly cached | $0.0096 | $0.030 |
| Cost per benchmark task (AA Index) | $0.602 | $2.35 |
| Output tokens per task (AA Index) | 139k | 113k |
| Of which reasoning | 91k | 80k |
| Time per task (AA Index) | 14 min | 14 min |
| Output speed | 172 tok/s | 132 tok/s |
| Time to first answer | 52.2 s | 25.9 s |
| Time to a complete answer | 102.8 s | 57.3 s |
Per token is not per task: an average task of Artificial Analysis' Intelligence Index cost $0.602 on DeepSeek V4.1 Flash 1017 (max) and $2.35 on Step 5 Preview, and DeepSeek V4.1 Flash 1017 (max) wrote 1.2× as many output tokens for the same work.
Each model is compared at its best-ranked reasoning-effort setting. DeepSeek V4.1 Flash 1017 (max) scores 59.2 and takes 52 s to its first answer token; DeepSeek V4.1 Flash 1017 (high) scores 51.7 in 24 s, DeepSeek V4.1 Flash 1017 (no reasoning) scores 45.2 in 1 s. Step 5 Preview scores 54.9 and takes 26 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
100 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 | DeepSeek V4.1 Flash 1017 (max) | Step 5 Preview |
|---|---|---|---|
| BenchLeader Index | BenchLeader | 59.2 | 54.9 |
| Agents & tools score | BenchLeader | 63.0 | 50.6 |
| Coding score | BenchLeader | 64.0 | 50.2 |
| Knowledge score | BenchLeader | 48.6 | 55.2 |
| Long context score | BenchLeader | 66.2 | 69.2 |
| Maths score | BenchLeader | 61.6 | 59.7 |
| Multimodal score | BenchLeader | 63.1 | 59.4 |
| Reasoning score | BenchLeader | 56.7 | 52.7 |
| Composite score | BenchLeader | – | 40.8 |
| Human preference score | BenchLeader | – | 63.2 |
| Blended price $/M | BenchLeader | $0.525 | $1.43 |
| Output speed | BenchLeader | 172 tok/s | 132 tok/s |
| Time to first answer | BenchLeader | 52.2 s | 25.9 s |
| Context window | BenchLeader | 1,048,576 tokens | 1,000,000 tokens |
| Reasoning | |||
| GPQA Diamond | Epoch AI Benchmarking Hub | 93.2% | – |
| Arena Hard Prompts | Arena (LMArena) | – | 1477 |
| LiveBench 2.0 Reasoning | LiveBench | 89.1%† | 83.8% |
| Humanity's Last Exam (AA) | Artificial Analysis | 43.3% | 46.5% |
| HLE-Diamond (AA) | Artificial Analysis | 25.3% | 24.3% |
| ARC-AGI-1 | ARC Prize | 93.5% | – |
| ARC-AGI-2 | ARC Prize | 78.1% | – |
| ARC-AGI-3 | ARC Prize | 5.8% | – |
| CritPt | Artificial Analysis | 24.0% | 20.9% |
| CritPt v2 hard subset (AA) | Artificial Analysis | 25.0% | 32.5% |
| MysteryMechanism | Vals AI | – | 18.5% |
| Chess Puzzles | Epoch AI Benchmarking Hub | 46.0% | – |
| Mystery Game Puzzles | Epoch AI Benchmarking Hub | 49.0% | – |
| Coding | |||
| SciCode | SciCode | 55.9% | 58.9% |
| Arena Coding | Arena (LMArena) | – | 1506 |
| Arena WebDev | Arena (LMArena) | 1665 | 1571† |
| LiveBench 2.0 Coding | LiveBench | 96.0%† | 85.6% |
| IOI | Vals AI | – | 38.1% |
| Code Migration | Vals AI | 44.2% | 39.0% |
| Code Migration v1.1 | Vals AI | 67.5% | 47.0% |
| SWE-bench Multimodal | Vals AI | – | 40.8% |
| Vibe Code Bench v1.1 | Vals AI | 87.7% | 70.4% |
| CursorBench | Cursor | 47.7% | – |
| ALE-Bench | Epoch AI Benchmarking Hub | 1622 | – |
| Terminal-Bench 4.0 (AA) | Artificial Analysis | 33.8% | 33.3% |
Show all 100 benchmarks
| Metric | Publisher | DeepSeek V4.1 Flash 1017 (max) | Step 5 Preview |
|---|---|---|---|
| Coding | |||
| Startrise Frontend | Startrise | 74.2%† | – |
| GameBench 2 Overall | DuelLab GameBench 2 | 61.3† | – |
| GameBench 2 by reasoning setting | DuelLab GameBench 2 | 69.8 | – |
| FrontierCode 1.1 (AA) | Artificial Analysis | 43.7% | 38.4% |
| FrontierCode 2.0 (AA) | Artificial Analysis | 28.5% | 12.0% |
| Agents & tools | |||
| APEX-Agents | Mercor | 69.8% | – |
| Arena Agent | Arena (LMArena) | 2.4 | -2.9 |
| LiveBench Long Agentic | LiveBench | 35.7%† | 19.4% |
| LiveBench Browser | LiveBench | 67.2%† | 24.8% |
| GDPval-AA v2.1, normalised score | Artificial Analysis | 62.3% | 54.3% |
| GDPval-AA v2.1 | Artificial Analysis | 1746 | 1594 |
| ITBench SRE (AA) | Artificial Analysis | – | 55.6% |
| Analyst Agent (AA) | Artificial Analysis | – | 35.0% |
| APEX-Agents (AA) | Artificial Analysis | – | 38.0% |
| Finance Agent v2 | Vals AI | 57.0% | 50.7% |
| Finance Agent v2.1 | Vals AI | 62.9% | 58.9% |
| Harvey's Legal Agent Benchmark v1.1 | Vals AI | 14.2% | 12.9% |
| Harvey's Legal Agent Benchmark v1 | Vals AI | 9.2% | 9.2% |
| Legal Research Bench | Vals AI | 45.2% | 40.4% |
| SREBench | Vals AI | – | 0.4% |
| Tax Agent Bench | Vals AI | 44.4% | 27.3% |
| Terminal-Bench 4.0 (Vals) | Vals AI | 31.8% | 31.8% |
| Terminal-Bench Science | Vals AI | – | 2.9% |
| AutomationBench | Artificial Analysis | 66.6% | 51.0% |
| GDP.pdf (AA) | Artificial Analysis | 17.2% | 14.8% |
| EnterpriseOps-Gym | Artificial Analysis | – | 47.2% |
| AA-Briefcase v1.1 | Artificial Analysis | 1570 | 1425 |
| GDPx | GDPx by SemiAnalysis | 5.4% | – |
| Terminal-Bench-Science 0.1 (AA) | Artificial Analysis | 12.9% | 2.4% |
| CWE-Bench (AA) | Artificial Analysis | 69.2% | 52.5% |
| DeepSec-Bench (AA) | Artificial Analysis | 24.9% | 15.8% |
| CyberGym end-to-end (AA) | Artificial Analysis | 76.3% | 74.8% |
| Maths | |||
| FrontierMath Tiers 1–3 | Epoch AI Benchmarking Hub | 71.2% | – |
| FrontierMath Tier 4 | Epoch AI Benchmarking Hub | 46.3% | – |
| OTIS Mock AIME | Epoch AI Benchmarking Hub | 99.4% | – |
| ProofBench | Vals AI | – | 42.0% |
| Arena Maths | Arena (LMArena) | – | 1466 |
| Knowledge | |||
| SimpleQA Verified | Epoch AI Benchmarking Hub | 42.5% | – |
| LiveBench 2.0 Data Analysis | LiveBench | 79.5%† | 63.5% |
| LiveBench Research | LiveBench | 79.5%† | 72.8% |
| AA-Omniscience v1.1 | Artificial Analysis | 0.9 | 16.4 |
| PRBench Finance | Scale AI SEAL | 47.1%† | – |
| PRBench Legal | Scale AI SEAL | 47.3%† | – |
| BioMysteryBench | Vals AI | – | 67.4% |
| Excel Modeling Benchmark | Vals AI | 61.8% | 57.1% |
| Excel Modeling Benchmark v1.1 | Vals AI | 69.2% | 69.5% |
| Public Benefits Bench v1.1 | Vals AI | – | 58.1% |
| DrugDiscoveryBench | Scale AI SEAL | 60.6%† | – |
| AA-Omniscience v1.1: accuracy | Artificial Analysis | 39.6% | 36.4% |
| AA-Omniscience v1.1: non-hallucination | Artificial Analysis | 35.9% | 68.7% |
| Instruction following | |||
| MultiChallenge | Scale AI SEAL | 74.9%† | – |
| Arena Instruction Following | Arena (LMArena) | – | 1446 |
| Human preference | |||
| Arena Text | Arena (LMArena) | – | 1448 |
| Arena Creative Writing | Arena (LMArena) | – | 1411 |
| Arena Multi-turn | Arena (LMArena) | – | 1442 |
| Arena Longer Queries | Arena (LMArena) | – | 1460 |
| Multimodal | |||
| Arena Vision | Arena (LMArena) | – | 1265 |
| LiveBench Design | LiveBench | 83.4%† | 76.9% |
| MMMU-Pro | Artificial Analysis | 80.6% | 76.4% |
| Code2Video-AA | Artificial Analysis | – | 821 |
| AA-3D | Artificial Analysis | – | 854 |
| Long context | |||
| AA-LCR v1.1 | Artificial Analysis | 82.7% | 88.3% |
| MLCR | Artificial Analysis | 26.7% | 18.3% |
| Composite | |||
| Epoch Capabilities Index | Epoch AI Benchmarking Hub | 159.3† | – |
| LiveBench 2.0 | LiveBench | 72.9%† | 58.6% |
| AA Intelligence Index v5.0 | Artificial Analysis | 32.6 | 28.5 |
| AA Intelligence Index v4.3.2 | Artificial Analysis | 45.1 | 43.7 |
| AA Progress Index v1 | Artificial Analysis | 323.7 | 318.2 |
| Vals Index | Vals AI | 61.3 | 54.8 |
| Vals Finance Index | Vals AI | 60.9% | 55.4% |
| Vals Science Index | Vals AI | – | 29.6% |
| Vals Router Index | Vals AI | 61.3%† | – |
| AA Cyber Index | Artificial Analysis | 56.8 | 47.7 |
| Safety & honesty | |||
| FORTRESS | Scale AI SEAL | 27.2%† | – |
| Metadata | |||
| AA Openness Index | Artificial Analysis | 44.4 | 33.3 |
Data as of 2026-11-06. Best configuration of each model; every score links to its source on the model pages.
DuelLab GameBench 2, release gb2-20261103-09b2218cbd85 (https://benchmarks.duellab.org/releases/gb2-20261103-09b2218cbd85/), CC BY 4.0 (https://creativecommons.org/licenses/by/4.0/). Scores mapped to BenchLeader model names; not endorsed by DuelLab.
What changed
- DeepSeek V4.1 Flash 1017 (max): 1 score revision.
- DeepSeek V4.1 Flash 1017 (max): 3 score revisions.
- DeepSeek V4.1 Flash 1017 (max): 1 new result (ARC-AGI-3); 2 score revisions.
- DeepSeek V4.1 Flash 1017 (max): 1 score revision.
- DeepSeek V4.1 Flash 1017 (max): 1 score revision.
- Step 5 Preview: 2 new results (Vals Index, Harvey's Legal Agent Benchmark v1.1).
- DeepSeek V4.1 Flash 1017 (max): 1 new result (Arena Agent); 1 score revision.
- DeepSeek V4.1 Flash 1017 (max): 4 new results (SciCode, APEX-Agents, CursorBench…); 1 score revision.
- Step 5 Preview: 1 score revision.
- DeepSeek V4.1 Flash 1017 (max): BenchLeader Index 61.8 → 59.0.
Related comparisons
- DeepSeek V4.1 Flash 1017 vs Claude Opus 5.5
- DeepSeek V4.1 Flash 1017 vs Gemini 4 Carbon
- DeepSeek V4.1 Flash 1017 vs GPT-6 Astra
- Step 5 Preview vs Gemini 4 Argon
- Step 5 Preview vs Gemini 3.8 Flash
- Step 5 Preview vs Muse Spark 1.3
- DeepSeek V4.1 Flash 1017 vs Claude Fable 5.1
- DeepSeek V4.1 Flash 1017 vs Claude Fable 5
Rankings: Best LLM overall · Best LLM for agents · Best cheap LLMs
DeepSeek V4.1 Flash 1017 vs Step 5 Preview: questions
- Is DeepSeek V4.1 Flash 1017 better than Step 5 Preview?
- DeepSeek V4.1 Flash 1017 (max) and Step 5 Preview are neck and neck on quality (59.2 vs 54.9, within the ±5.7 margin). The BenchLeader Index combines every independent quality benchmark; the two are within the margin of error as of 6 Nov 2026, so check the category scores for your use.
- DeepSeek V4.1 Flash 1017 vs Step 5 Preview: which is cheaper, and will that change?
- DeepSeek V4.1 Flash 1017 is cheaper today: $0.525 against $1.43 per million tokens, blended at three input tokens per output token. No price change has been announced for either.
- Is DeepSeek V4.1 Flash 1017 or Step 5 Preview better for coding?
- DeepSeek V4.1 Flash 1017 scores higher in coding (64 vs 50 on the category index, where 50 is average). On FrontierCode 2.0 (AA), the top shared coding benchmark, it is 28.5% for DeepSeek V4.1 Flash 1017 against 12.0% for Step 5 Preview.
- Which is faster, DeepSeek V4.1 Flash 1017 or Step 5 Preview?
- DeepSeek V4.1 Flash 1017 streams faster (172 vs 132 output tokens per second, Artificial Analysis 7-day median).
- Is DeepSeek V4.1 Flash the same as DeepSeek V4.1 Flash 1017?
- DeepSeek V4.1 Flash 1017 is the checkpoint compared here (BenchLeader Index 59.2); people searching for DeepSeek V4.1 Flash land on this page because it replaced that earlier build.
More questions (5)
- Is DeepSeek V4.1 Flash 1017 or Step 5 Preview better for agents and tool use?
- DeepSeek V4.1 Flash 1017 scores higher in agentic tasks (63 vs 51 on the category index, where 50 is average). On Arena Agent, the top shared agents & tools benchmark, it is 2.4 for DeepSeek V4.1 Flash 1017 against -2.9 for Step 5 Preview.
- Which is better for long documents?
- Step 5 Preview scores higher on long-context benchmarks (69 vs 66). Both take about 1M tokens: DeepSeek V4.1 Flash 1017 1,048,576, Step 5 Preview 1,000,000.
- Which has lower latency?
- Step 5 Preview answers first: 25.9 s against 52.2 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 DeepSeek V4.1 Flash 1017 vs Step 5 Preview?
- Across the 40 benchmarks both have, DeepSeek V4.1 Flash 1017 leads 31, Step 5 Preview leads 7 and 2 are tied. Every overall index that rates both puts DeepSeek V4.1 Flash 1017 ahead.
- When was DeepSeek V4.1 Flash 1017 released?
- DeepSeek V4.1 Flash 1017 was released on 17 Oct 2026.