What are DeepSeek V4 Pro's benchmark scores?

As of 24 Oct 2026, DeepSeek V4 Pro (max effort) scores 60.8 ±3.1 on the BenchLeader Index, strongest in composite (70) and weakest in maths (56). That rests on 26 independent benchmark results, and every score on this page links to its publisher.

How does DeepSeek V4 Pro rank?

DeepSeek V4 Pro (max effort) is #103 of 785 ranked configurations on 24 Oct 2026. 198 other configurations score within its margin, so the exact position matters less than the band. The leader is Claude Opus 5.5 (max effort) at 72.6.

How much does DeepSeek V4 Pro cost per 1M tokens?

As of 24 Oct 2026, DeepSeek V4 Pro costs $1.32 per 1M input tokens and $3.96 per 1M output tokens, $1.98 blended (three input tokens per output token). The list price last changed on 6 Oct 2026.

Is DeepSeek V4 Pro better than GPT-5.6 Sol?

As of 24 Oct 2026, GPT-5.6 Sol (max) leads on quality (68.0 vs 60.8 on the BenchLeader Index), but DeepSeek V4 Pro (max) costs 4.0× less, so it is the better buy unless you need the extra quality. The head-to-head page has the benchmark-by-benchmark detail.

When was DeepSeek V4 Pro released?

DeepSeek released DeepSeek V4 Pro on 13 Aug 2026. BenchLeader has results for it at 6 reasoning-effort settings.

DeepSeek Open weights: deepseek-ai/DeepSeek-V4-Pro Compare with another model

DeepSeek V4 Pro

Reasoning effort

DeepSeek V4 Pro is a DeepSeek open-weights model, released 13 Aug 2026. Its best configuration (max reasoning effort) ranks #103 of 785 on the BenchLeader Index at 60.8 ±3.1, in the upper half. The ± is the point: 198 other configurations score within that range, so they and this one cannot be told apart on quality alone — price and speed are what separate them. It scores highest in composite (70) and lowest in maths (56). At $1.98 per million tokens blended it is pricier than most ranked models. Output speed of 93 tokens per second puts it faster than most, with a first answer in 23.3 s. It has been measured at 6 reasoning-effort settings; this summary describes the best-scoring one, and the tabs above switch between them. Last measured 20 Oct 2026.

Blended price: $1.98/M; $1.32 in · $3.96 out
Output speed: 93 tok/s; measured by Artificial Analysis
First answer: 23 s; first token 1.61 s
Context: 1.0M
Full answer: 29 s; median, reasoning included
Cost per run: $0.674; one full Intelligence Index run
Released: 13 Aug 2026

How it scores by categoryDashed line = average model (50). One step of 15 = one standard deviation.

Overall index61
Reasoning65
Coding58
Agents & tools60
Maths56
Knowledge60
Long context65
Composite70

Versions

DeepSeek has shipped 2 models under this name. Each is ranked on its own results; a newer version often has fewer results so far, which holds its index nearer the average until more arrive.

Model	Released	Index	Rank
DeepSeek V4 Promaxthis page	13 Aug 2026	60.8	#103
DeepSeek V4 Pro 0424max	24 Apr 2026	59.6	#128

Reasoning-effort configurations

The same model behaves differently depending on how much it is allowed to think. Each row is one setting, scored only on the benchmarks that were run at that setting. “Not stated” collects results from publishers that did not say which setting they used; for a reasoning model that is usually its thinking mode, but we do not assume it. Pick a setting here or at the top of the page to see its price, speed and category scores.

Effort	Index	Rank	Speed	First answer	Chat reply cost	Categories
no reasoning	46.9	#469	96 tok/s	1.71 s	$0.0017	Coding 34 · Composite 52 · Knowledge 56 · Long context 50 · Maths 44 · Reasoning 40
low	–	–	50 tok/s	1.61 s	$0.0009	Reasoning 60
high	57.5	#176	50 tok/s	1.61 s	$0.0009	Agents & tools 49 · Coding 54 · Human preference 65 · Maths 65 · Reasoning 63
xhigh	–	–	50 tok/s	1.61 s	$0.0009
maxbest	60.8	#103	93 tok/s	23 s	$0.0017	Agents & tools 60 · Coding 58 · Composite 70 · Knowledge 60 · Long context 65 · Maths 56 · Reasoning 65
not stated	53.5	#285	50 tok/s	1.61 s	$0.0009	Agents & tools 44 · Coding 53 · Human preference 58 · Maths 60 · Reasoning 57

How its index has moved

8 Sept 2026 to 21 Oct 2026

max
high
not stated
no reasoning

The index is recomputed from scratch every day, so a line moves when a new benchmark result lands, when a publisher revises a score, or when the models it is normalised against change. Early movement usually means the score is still settling.

Benchmark results

See all LLM benchmarks. One column per reasoning effort. Rank is among every configuration of every model on that benchmark. Hover a score for the run it came from.

Metadata

Benchmark	max	Source	Trend
AA Openness Indexnot in index	44.4#27	Artificial Analysis

Reasoning

Benchmark	no reasoning	low	high	max	not stated	Source
GPQA Diamond	73.2%#158	–	90.9%#35	91.7%#30	–	Epoch AI Benchmarking Hub
Arena Hard Prompts	–	–	1483#61	–	1480#65	Arena (LMArena)
LiveBench 1.0 Reasoningnot in indexeffort not stated	–	–	–	–	85.8%#34	LiveBench
GPQA Diamond (AA)not in index	–	–	–	92.8%#28	–	Artificial Analysis
Humanity's Last Exam (AA)not in index	10.6%#314	–	–	41.0%#81	–	Artificial Analysis
GPQA Diamond (Vals)not in index	–	–	–	92.4%#16	–	Vals AI
Kagi LLM Benchmarkeffort not stated	–	–	–	–	53.5%#70	Kagi LLM Benchmark
ARC-AGI-1	13.0%#231	90.5%#74	87.2%#97	90.0%#80	–	ARC Prize
ARC-AGI-2	0.8%#231	56.3%#98	59.7%#89	61.3%#82	–	ARC Prize
CritPt	0.0%#315	–	–	18.0%#72	–	Artificial Analysis
Chess Puzzlesnot in index	4.0%#133	–	15.0%#107	46.0%#26	–	Epoch AI Benchmarking Hub
Mystery Game Puzzlesnot in index	17.0%#75	–	–	43.0%#16	–	Epoch AI Benchmarking Hub
CL-bench Lifenot in index	–	–	13.5%#9	–	–	Epoch AI Benchmarking Hub
LMCAnot in index	–	–	–	41.2%#105	45.5%#88	Epoch AI Benchmarking Hub
DTBenchnot in index	–	–	–	90.7%#79	93.9%#60	Epoch AI Benchmarking Hub
ForecastBenchnot in indexeffort not stated	–	–	–	–	60.9%#70	ForecastBench

Coding

Benchmark	no reasoning	high	max	not stated	Source
SWE-bench Verified (Epoch)	–	–	77.6%#6	–	Epoch AI Benchmarking Hub
SciCode	39.9%#166	46.4%#133	51.0%#100	–	SciCode
WeirdML	–	46.5%#78	66.2%#39	–	WeirdML
FrontierCode	17.6%#40	28.6%#33	–	–	Cognition
Arena Coding	–	1505#68	–	1501#78	Arena (LMArena)
Arena WebDev	–	1582#34	–	1447#72	Arena (LMArena)
LiveBench 1.0 Codingnot in indexeffort not stated	–	–	–	77.2%#39	LiveBench
SciCode (AA)not in index	39.9%#177	–	51.0%#100	–	Artificial Analysis
LiveCodeBench	–	–	87.5%#13	–	Vals AI
IOI	–	–	51.6%#24	–	Vals AI
SWE-bench (Vals)not in index	–	–	96.4%#2	–	Vals AI
Code Migrationnot in index	–	–	41.5%#28	–	Vals AI
Code Migration v1.1not in index	–	–	52.0%#14	–	Vals AI
ProgramBenchnot in index	–	–	0.0%#32	–	Vals AI
Vibe Code Bench 1-100not in index	–	–	17.5%#16	–	Vals AI
Vibe Code Bench v1.1not in index	–	–	82.3%#22	–	Vals AI
SWE Atlas: Codebase QnAnot in indexeffort not stated	–	–	–	27.1%#16	Scale AI SEAL
SWE Atlas: Test Writingnot in indexeffort not stated	–	–	–	27.1%#20	Scale AI SEAL
Surface Evolver Benchnot in index	–	40.0%#23	–	–	Epoch AI Benchmarking Hub
ALE-Benchnot in index	521.7#103	1006.1#57	1403.2#27	–	Epoch AI Benchmarking Hub
Terminal-Bench 4.0 (AA)not in index	1.5%#122	–	14.1%#74	–	Artificial Analysis
Terminal-Bench 2.1 (AA)not in index	–	–	78.7%#49	–	Artificial Analysis
Startrise Frontendnot in indexeffort not stated	–	–	–	62.7%#28	Startrise
FrontierCode 1.1 (AA)not in index	–	–	38.8%#36	–	Artificial Analysis

Agents & tools

Benchmark	no reasoning	high	max	not stated	Source
APEX-Agentseffort not stated	–	–	–	47.3%#33	Mercor
Arena Agent	–	0.1#29	–	–	Arena (LMArena)
LiveBench Agentic Codingnot in indexeffort not stated	–	–	–	55.0%#30	LiveBench
Terminal-Bench 2.1 (Vals)	–	–	54.7%#49	50.2%#57	Vals AI
HiL-Bencheffort not stated	–	–	–	38.5%#19	Scale AI SEAL
GDPval-AA v2.1not in index	27.8%#89	–	47.5%#43	–	Artificial Analysis
τ³-Banking (AA)	–	–	39.6%#31	–	Artificial Analysis
CyberBenchnot in index	–	–	72.9%#14	–	Vals AI
Finance Agent v2not in index	–	–	50.4%#44	–	Vals AI
Finance Agent v2.1not in index	–	–	57.1%#23	–	Vals AI
Harvey's Legal Agent Benchmark v1.1not in index	–	–	10.8%#14	–	Vals AI
Harvey's Legal Agent Benchmark v1not in index	–	–	7.5%#28	–	Vals AI
Legal Research Benchnot in index	–	–	40.9%#28	–	Vals AI
SkillsBenchnot in index	–	–	53.8%#16	–	Vals AI
SREBenchnot in index	–	–	1.9%#24	–	Vals AI
Tax Agent Benchnot in index	–	–	26.8%#34	–	Vals AI
Terminal-Bench 2.0 (Vals)not in indexeffort not stated	–	–	–	56.2%#14	Vals AI
Terminal-Bench 4.0 (Vals)not in index	–	–	14.1%#34	–	Vals AI
Terminal-Bench Sciencenot in index	–	–	4.3%#25	–	Vals AI
Vending-Bench 2not in indexeffort not stated	–	–	–	3284.5#48	Andon Labs
AutomationBenchnot in index	34.9%#110	–	56.7%#52	–	Artificial Analysis
GDP.pdf (AA)not in index	3.6%#177	–	11.4%#119	–	Artificial Analysis
EnterpriseOps-Gymnot in index	–	–	49.6%#4	–	Artificial Analysis
AA-Briefcase v1.1not in index	–	–	1256#44	–	Artificial Analysis
Terminal-Bench-Science 0.1 (AA)not in index	–	–	5.7%#44	–	Artificial Analysis

Maths

Benchmark	no reasoning	high	max	not stated	Source
FrontierMath Tiers 1–3	–	–	64.6%#37	–	Epoch AI Benchmarking Hub
FrontierMath Tier 4	–	–	26.8%#43	–	Epoch AI Benchmarking Hub
OTIS Mock AIME	46.7%#187	95.6%#44	98.6%#23	–	Epoch AI Benchmarking Hub
ProofBench	–	–	16.0%#56	50.0%#30	Vals AI
LiveBench Mathematicsnot in indexeffort not stated	–	–	–	95.1%#17	LiveBench
AIME 2026possible contamination	–	–	96.7%#6	–	MathArena
HMMT February 2026possible contamination	–	–	93.9%#10	–	MathArena
USAMO 2026not in indexpossible contamination	–	–	60.7%#4	–	MathArena
MathArena Apexpossible contamination	–	–	28.1%#10	–	MathArena
Arena Maths	–	1466#58	–	1444#84	Arena (LMArena)

Knowledge

Benchmark	no reasoning	xhigh	max	not stated	Source
SimpleQA Verified	–	–	52.9%#23	–	Epoch AI Benchmarking Hub
LiveBench 1.0 Data Analysisnot in indexeffort not stated	–	–	–	79.2%#18	LiveBench
AA-Omniscience	-12.4#211	–	0.8#129	–	Artificial Analysis
MMLU-Pro	–	–	87.3%#32	–	Vals AI
LegalBench	–	–	82.4%#71	–	Vals AI
CorpFin	–	–	65.4%#28	–	Vals AI
TaxEval	–	–	73.1%#52	–	Vals AI
CaseLaw v2not in index	–	–	59.4%#27	–	Vals AI
Excel Modeling Benchmarknot in index	–	–	52.8%#52	–	Vals AI
Excel Modeling Benchmark v1.1not in index	–	–	65.1%#25	–	Vals AI
MedCodenot in index	–	–	42.5%#49	–	Vals AI
MedScribenot in index	–	–	80.2%#57	–	Vals AI
Public Benefits Bench v1.1not in index	–	–	62.9%#25	–	Vals AI
DrugDiscoveryBenchnot in index	–	31.7%#14	–	–	Scale AI SEAL
AA-Omniscience: accuracynot in index	41.5%#113	–	49.1%#67	–	Artificial Analysis
AA-Omniscience: non-hallucinationnot in index	8.1%#470	–	5.2%#521	–	Artificial Analysis

Instruction following

Benchmark	high	not stated	Source
LiveBench Languagenot in indexeffort not stated	–	82.1%#26	LiveBench
LiveBench Instruction Followingnot in indexeffort not stated	–	67.7%#35	LiveBench
Arena Instruction Followingnot in index	1461#51	1453#60	Arena (LMArena)

Human preference

Benchmark	high	not stated	Source
Arena Text	1465#57	1458#64	Arena (LMArena)
EQ-Bench 4effort not stated	–	1166#17	EQ-Bench
Arena Creative Writingnot in index	1446#48	1446#45	Arena (LMArena)
Arena Multi-turnnot in index	1471#61	1475#50	Arena (LMArena)
Arena Longer Queriesnot in index	1475#56	1472#60	Arena (LMArena)

Long context

Benchmark	no reasoning	max	Source	Trend
AA-LCR	50.7%#320	80.3%#71	Artificial Analysis

Composite

Benchmark	no reasoning	max	not stated	Source
Epoch Capabilities Indexnot in indexeffort not stated	–	–	155.6#34	Epoch AI Benchmarking Hub
LiveBench 1.0not in indexeffort not stated	–	–	77.4%#23	LiveBench
AA Intelligence Index v4.3.2	20.4#232	36#86	–	Artificial Analysis
Vals Indexnot in index	–	53.0#17	–	Vals AI
Vals Finance Indexnot in index	–	52.9%#24	–	Vals AI
Vals Router Indexnot in indexeffort not stated	–	–	53.0%#16	Vals AI

Where to run it

Every provider serving this model through OpenRouter, with throughput and first-token latency measured on live traffic over the last 30 minutes and each provider’s own price. Purple marks the best in each column.

Provider	Speed	First token	Input $/M	Output $/M	Context	Quantisation
Azure	99 tok/s	1.70 s	$1.91	$3.83	1.0M	–
Relace	96 tok/s	0.68 s	$0.900	$1.50	1.0M	fp4
NovitaAI	65 tok/s	2.05 s	$1.60	$3.20	1.0M	fp8
DeepInfra	64 tok/s	0.94 s	$1.30	$2.60	1.0M	fp8
Baidu Qianfan	61 tok/s	0.83 s	$0.671	$1.34	1.0M	fp8
DigitalOcean	61 tok/s	0.74 s	$1.04	$2.09	1.0M	–
Cloudflare	50 tok/s	1.07 s	$1.15	$2.55	1.0M	–
Reka AI	48 tok/s	0.81 s	$1.05	$2.10	1.0M	–
SiliconFlow	42 tok/s	1.28 s	$1.50	$3.13	1.0M	fp8
Venice	42 tok/s	1.66 s	$1.65	$3.30	1M	–
StreamLake	41 tok/s	1.42 s	$0.662	$1.32	1.0M	fp8
AtlasCloud	39 tok/s	1.78 s	$1.68	$3.38	1.0M	fp4
Alibaba Cloud Int.	39 tok/s	1.81 s	$1.42	$2.83	1M	fp8
Alibaba Cloud Int.	35 tok/s	1.76 s	$1.56	$3.12	1M	fp8
GMICloud	26 tok/s	1.96 s	$0.957	$1.91	1.0M	fp8

Price history

The price above, per 1M tokens, every day it changed.

input outputnow $1.32 in · $3.96 out

What a task costs

Estimates from list price, output speed and time to first answer for the best configuration. “With caching” assumes three-quarters of the input is served from the prompt cache at $0.022 per 1M. Reasoning tokens are not modelled.

Workload	Tokens in / out	Cost	With caching	Time
Chat reply	400 / 300	$0.0017	$0.0013	26.5 s
Summarize a 30-page report	12,000 / 600	$0.018	$0.0065	29.7 s
Code edit	6,000 / 1,500	$0.014	$0.0080	39.5 s
Agentic coding session	60,000 / 4,000	$0.095	$0.037	1.1 min
Structured extraction	2,000 / 200	$0.0034	$0.0015	25.4 s

New this month

Compare DeepSeek V4 Pro with…

Questions people ask

What are DeepSeek V4 Pro's benchmark scores?: As of 24 Oct 2026, DeepSeek V4 Pro (max effort) scores 60.8 ±3.1 on the BenchLeader Index, strongest in composite (70) and weakest in maths (56). That rests on 26 independent benchmark results, and every score on this page links to its publisher.
How does DeepSeek V4 Pro rank?: DeepSeek V4 Pro (max effort) is #103 of 785 ranked configurations on 24 Oct 2026. 198 other configurations score within its margin, so the exact position matters less than the band. The leader is Claude Opus 5.5 (max effort) at 72.6.
How much does DeepSeek V4 Pro cost per 1M tokens?: As of 24 Oct 2026, DeepSeek V4 Pro costs $1.32 per 1M input tokens and $3.96 per 1M output tokens, $1.98 blended (three input tokens per output token). The list price last changed on 6 Oct 2026.
Is DeepSeek V4 Pro better than GPT-5.6 Sol?: As of 24 Oct 2026, GPT-5.6 Sol (max) leads on quality (68.0 vs 60.8 on the BenchLeader Index), but DeepSeek V4 Pro (max) costs 4.0× less, so it is the better buy unless you need the extra quality. The head-to-head page has the benchmark-by-benchmark detail.
When was DeepSeek V4 Pro released?: DeepSeek released DeepSeek V4 Pro on 13 Aug 2026. BenchLeader has results for it at 6 reasoning-effort settings.

Data as of 24 Oct 2026. Compare these configurations.

Cite as: BenchLeader, “DeepSeek V4 Pro: benchmarks, pricing, speed and rank”, data as of 24 Oct 2026, https://www.benchleader.com/models/deepseek-v4-pro.