OpenAIDeepSeekLLM matchup

GPT-6 Lunavs.DeepSeek-V4-Pro

Which one is better?

GPT-6 Luna performs better than DeepSeek-V4-Pro and is cheaper: DeepSeek-V4-Pro costs 6.6 times (input) and 4 times (output) as much.

GPT-6 Luna ahead

IntelligenceOverall index, 100 points = average

96.3

GPT-6 Luna

Rank 35

86.8

DeepSeek-V4-Pro

Rank 43

GPT-6 Luna faster

SpeedTokens per second

101

GPT-6 Luna

DeepSeek-V4-Pro

GPT-6 Luna cheaper

Price$ per 1M tokens

$0.20

GPT-6 Luna

$0.99

DeepSeek-V4-Pro

about as popular

Popularity0 to 100

GPT-6 Luna

Rank 30 of 785

DeepSeek-V4-Pro

Rank 49 of 785

Above 272,000 tokens of context, GPT-6 Luna costs more: $0.20 input and $0.75 output per 1M tokens.

Choose GPT-6 Luna for

MathematicsHealthReasoningFinanceGeneralCodingAgentsLegal (not certain)

Choose DeepSeek-V4-Pro for

No clear advantage

Where is the difference?

Strengths at a glance

GPT-6 LunaDeepSeek-V4-ProMedian of all ranked models

CodingGPT-6 Luna ahead

GPT-6 Luna97.5Rank 35

DeepSeek-V4-Pro90.4Rank 49

Median of ranked models84.1

95 % range: GPT-6 Luna 94.5 to 100.5, DeepSeek-V4-Pro 85.8 to 95.0.

AreaGPT-6 LunaDeepSeek-V4-Pro

The same values and verdicts as the bars below. In the close-up the net starts at 75 points, just below the weaker model's lowest value; the reference group average is 100. Not in the net because a model has no own value there or the data are too thin: Multimodal.

GPT-6 Luna is ahead in 7 of 10 areas, DeepSeek-V4-Pro in none.

GPT-6 Luna ahead: 7GPT-6 Luna slightly ahead: 1level: 1too little data: 1

Areas in the overall index

MathematicsGPT-6 Luna +20105.5 to 85.4

ReasoningGPT-6 Luna +1397.8 to 84.6

GeneralGPT-6 Luna +794.4 to 87.1

CodingGPT-6 Luna +797.5 to 90.4

AgentsGPT-6 Luna +799.5 to 93.0

Knowledgelevel89.2 to 88.3

Multimodalonly GPT-6 Luna has a score

Specialist fields

HealthGPT-6 Luna +1395.2 to 81.8

FinanceGPT-6 Luna +1198.3 to 87.1

Legalleaning GPT-6 Lunanot certain · 3 and 2 tests

Bar length shows the gap in index points (full width 15 points, one standard deviation of the reference group). Solid means clearly ahead, hatched not certain yet, a thin grey bar level.

Show 4 more areas that do not count toward the overall index

Productivitytoo little data90.5 to 89.5

Cybersecurityonly GPT-6 Luna has a score

Searchonly GPT-6 Luna has a score

Educationonly GPT-6 Luna has a score

Bar length shows the gap in index points (full width 15 points, one standard deviation of the reference group). Solid means clearly ahead, hatched not certain yet, a thin grey bar level.

GPT-6 Luna vs. DeepSeek-V4-Pro at the same effort

Both models measured at the same reasoning efforts. This shows whether a lead comes from the model or only from more compute.

noneDeepSeek-V4-Pro slightly ahead

GPT-6 Luna 59.4 vs. DeepSeek-V4-Pro 62.5 · Gap 6.6 · Test series 10

maxGPT-6 Luna ahead

GPT-6 Luna 96.3 vs. DeepSeek-V4-Pro 86.8 · Gap 5.3 · Test series 11

Overall index (max vs. max)GPT-6 Luna ahead

GPT-6 Luna 96.3 vs. DeepSeek-V4-Pro 86.8

Effort	GPT-6 Luna	DeepSeek-V4-Pro	Gap	Verdict	Test series
none	59.4	62.5	6.6	DeepSeek-V4-Pro slightly ahead	10
max	96.3	86.8	5.3	GPT-6 Luna ahead	11
Overall index (max vs. max)	96.3	86.8	9.4	GPT-6 Luna ahead

Gap and verdict compare only the Artificial Analysis test series in which both models are measured at that effort. Tied below 3 points, ahead only beyond two standard errors. The single values rest on all test series of each model, so their difference can deviate from the gap.

Which model for which task?

Everyday tasks, translated from the area values above. Each row names the values it rests on.

TaskBetter choicebased on

Writing and debugging codeGPT-6 LunaCoding 97.5 to 90.4

Maths and data analysisGPT-6 LunaMathematics 105.5 to 85.4, Reasoning 97.8 to 84.6

Facts and researchlevelKnowledge 89.2 to 88.3, Search without a score for DeepSeek-V4-Pro

Agents and workflowsGPT-6 LunaAgents 99.5 to 93.0

Fast chat repliesGPT-6 Luna101 to 70 tokens per second

Security analysisopenCybersecurity without a score for DeepSeek-V4-Pro

Many requests on a small budgetGPT-6 Luna$0.20 to $0.99 per 1M tokens

Finance, legal and medical textsGPT-6 LunaFinance 98.3 to 87.1, Legal 90.7 to 83.7, Health 95.2 to 81.8

Images, screenshots and PDFsopenMultimodal without a score for DeepSeek-V4-Pro

See for yourself: same task, one attempt

The same prompt for both models. We show the latest attempt without fixes; it does not count toward the overall index.

Pelican on a bicycle

Prompt

Generate an SVG of a pelican riding a bicycle

Task origin: Simon Willison

GPT-6 Luna

Oct 1, 2026 · 47 s · $0.0022

Test details

Run date: Oct 1, 2026
API provider: openrouter
Requested API model: openai/gpt-6-luna
Reasoning setting: high
Token limit including reasoning: 8,192 tokens
Output tokens: 4,493 tokens
Cost of the attempt: $0.0022
Response time: 46.7 s
API model reported in response: openai/gpt-6-luna
Pinned endpoint: OpenAI | openai/gpt-6-luna-20260922
Temperature: Provider default
Test protocol: community-visual-tests-v1
Technical checks: Passed, not a quality rating

Visually reviewed. This individual test does not establish overall model quality.

DeepSeek-V4-Pro

The response reached the token limit. No complete recording is available.

Oct 5, 2026 · 68 s · $0.02

Not yet visually reviewed. A successful recording does not confirm correct physics or full compliance with the prompt.

Test details

Run date: Oct 5, 2026
API provider: openrouter
Requested API model: deepseek/deepseek-v4-pro-0813
Reasoning setting: high
Token limit including reasoning: 8,192 tokens
Output tokens: 8,192 tokens
Cost of the attempt: $0.0163
Response time: 67.6 s
API model reported in response: deepseek/deepseek-v4-pro-0813
Pinned endpoint: DeepSeek | deepseek/deepseek-v4-pro-20260813
Temperature: Provider default
Test protocol: community-visual-tests-v1
Technical checks: Failed

Earlier attempts (3)

Oct 2, 2026API request failedReasoning setting: high
Oct 1, 2026API request failedReasoning setting: high
Sep 7, 2026API request failedReasoning setting: high

SVG short test

One SVG task with the same fixed settings for every model, without hidden thinking.

Prompt

Return only one standalone SVG of a pelican riding a bicycle. Use viewBox="0 0 800 600", simple shapes, two clearly visible wheels and a recognizable pelican with a long beak. At most 30 SVG elements and 1800 characters. No Markdown, explanations, comments, scripts, animation, CSS or external resources.

GPT-6 Luna

Unavailable in this round

DeepSeek-V4-Pro

GPT-6 Luna and DeepSeek-V4-Pro in the overall index

Every model with a place in the overall index, both models of this matchup highlighted. You can add or hide other models.

Overall index (Gradually Meta-Score v4)

20 of 92 models

4 models in this view are measured at a lower reasoning effort than max in the overall index. Dashed column: lower reasoning effort than the top.

Claude Opus 5.5max126.1

Gemini 4 Carbonhigh125.8

Claude Fable 5.1max123.8

Claude Sonnet 5.5max122.8

Gemini 4 Argonhigh122.3

GPT-6 Astramax121.3

GPT-6.1 Solmax119.2

Muse Wonder Fullmax118.3

Claude Opus 5max117.2

Muse Wondermax113.4

Kimi K3.1 Promax112.9

Claude Fable 5max112.5

GPT-6 Solmax110.5

GPT-5.6 Solmax108.7

Muse Spark 1.3max107.4

Gemini 3.7 Flashhigh107.1

DeepSeek V4.1 Flash 1017max106.6

Gemini 3.8 Flashhigh106.5

GPT-6 Lunamax96.3

DeepSeek-V4-Promax86.8

Claude Opus 5.5

Gemini 4 Carbon

Claude Fable 5.1

Claude Sonnet 5.5

Gemini 4 Argon

GPT-6 Astra

GPT-6.1 Sol

Muse Wonder Full

Claude Opus 5

Muse Wonder

Kimi K3.1 Pro

Claude Fable 5

GPT-6 Sol

GPT-5.6 Sol

Muse Spark 1.3

Gemini 3.7 Flash

DeepSeek V4.1 Flash 1017

Gemini 3.8 Flash

GPT-6 Luna

DeepSeek-V4-Pro

Popularity compared

How often the models are searched for, used and mentioned, on a scale from 0 to 100.

Both are about as popular: GPT-6 Luna 60, DeepSeek-V4-Pro 56 out of 100.

Over time

Weekly values from October 18, 2026 to November 7, 2026

GPT-6 LunaDeepSeek-V4-Pro

All data side by side

The better value is bold. For the overall index only when the gap is clear.

Metric	GPT-6 Luna	DeepSeek-V4-Pro
Intelligence
Overall indexpoints, 100 = reference group average, higher is better	96.3	86.8
Rank	Rank 35 of 92	Rank 43 of 92
Price per 1M tokens
Inputwhat you send	$0.10above 272,000 tokens: $0.20	$0.66
Outputwhat the model writes	$0.50above 272,000 tokens: $0.75	$1.98
Cost per task (own tests)average of 1 original-protocol task both models passed (SVG short test not included)	$0.003	$0.021
Speed
Tokens per second	101	70
Popularity
Popularitysearch, usage and mentions, 0 to 100	60Rank 30 of 785	56Rank 49 of 785
Technical
Context windowhow much text the model reads at once	1.05 M tokens≈ 1,580 A4 pages	1 M tokens≈ 1,500 A4 pages
Knowledge cutoff	May 18, 2026	n/a
Released	Sep 22, 2026	Apr 24, 2026
Availability	Active	Active
Open weights	no	yes

DeepSeek-V4-Pro: At peak times (Mon, Tue, Wed, Thu, Fri, 01:00-04:00 UTC; Mon, Tue, Wed, Thu, Fri, 06:00-10:00 UTC) the price is $1.32 input and $3.96 output per 1M tokens. The table shows the off-peak price.

Prices without caching or batch discounts · Page count as a rule of thumb from Artificial Analysis

Is there a better choice?

As cheap as GPT-6 LunaClaude Haiku 5.5

Overall index 103.4, rank 20$0.10 input and $0.50 output per 1M tokens

GPT-6 Luna vs. Claude Haiku 5.5

Best model in the overall indexClaude Opus 5.5

Overall index 126.1, rank 1$4 input and $20 output per 1M tokens

GPT-6 Luna vs. Claude Opus 5.5

Shared benchmarks

Single results from the same benchmark version, task and metric. A value is bold only with the same test setup and a clear gap.

A: GPT-6 LunaB: DeepSeek-V4-Pro

AA-BriefcaseOverall · higher is better1,337.1 elo1,253.2 eloArtificial AnalysisAs of 11/07/2026

AA-OmniscienceOverall · higher is better-8.267 points-6.367 pointsArtificial AnalysisAs of 11/07/2026

AA-Omniscience AccuracyOverall · higher is better39.6%45.2%Artificial AnalysisAs of 11/07/2026

AA-Omniscience Hallucination RateOverall · lower is better79.3%94.1%Artificial AnalysisAs of 11/07/2026

AutomationBench-AAOverall · higher is better53.2%56.7%Artificial AnalysisAs of 11/07/2026

Code MigrationOverall · higher is betterTest setup not documented as identical42.6%26.2%Vals AIAs of 10/29/2026

CritPt v2 (Hard Subset)Overall · higher is better42.5%23.5%Artificial AnalysisAs of 11/07/2026

EMBOverall · higher is betterTest setup not documented as identical68.5%51.6%Vals AIAs of 10/20/2026

Finance Agent (v2)Overall · higher is betterTest setup not documented as identical49.9%44.1%Vals AIAs of 10/29/2026

FrontierCodeOverall · higher is better25.2%17.1%Artificial AnalysisAs of 11/07/2026

Show 54 more benchmarks

FrontierCode 1.1 MainOverall · higher is better41.9%38.8%Artificial AnalysisAs of 11/07/2026

GDP.pdfOverall · higher is better22.8%11.4%Artificial AnalysisAs of 11/07/2026

GDPval-AAOverall · higher is better1,444.5 elo1,457.4 eloArtificial AnalysisAs of 11/07/2026

HLE-DiamondOverall · higher is better22.2%19.5%Artificial AnalysisAs of 11/07/2026

Legal Research BenchOverall · All-pass · higher is betterTest setup not documented as identical30.3%23.1%Vals AIAs of 11/05/2026

LMArena Code Arena WebDevOverall · higher is betterTest setup not documented as identical1,5791,464LMArenaAs of 10/28/2026

LMArena Text, Style ControlOverall · higher is betterTest setup not documented as identical1,4411,455LMArenaAs of 10/28/2026

MedCodeOverall · higher is better44.7%40.5%Vals AIAs of 10/29/2026

MedScribeOverall · higher is better83.7%75.1%Vals AIAs of 10/29/2026

ProgramBenchOverall · higher is betterTest setup not documented as identical0.5%0.0%Vals AIAs of 11/05/2026

ProofBench v1.1Overall · higher is better64%16%Vals AIAs of 10/29/2026

Public Benefits Bench v1.1Overall · higher is betterTest setup not documented as identical57.6%62.9%Vals AIAs of 10/12/2026

Tax Agent BenchOverall · higher is betterTest setup not documented as identical19.9%25.2%Vals AIAs of 11/05/2026

Terminal-BenchOverall · higher is better12.6%14.1%Artificial AnalysisAs of 11/07/2026

Terminal-Bench 4.0Overall · higher is betterTest setup not documented as identical13.6%11.1%Vals AIAs of 11/05/2026

Terminal-Bench-ScienceOverall · higher is better8.6%5.7%Artificial AnalysisAs of 11/07/2026

Vibe Code Bench v1.1Overall · higher is betterTest setup not documented as identical81.6%49.9%Vals AIAs of 11/05/2026

Code MigrationCLI · higher is betterTest setup not documented as identical37.2%15.5%Vals AIAs of 10/29/2026

EMBComps · higher is betterTest setup not documented as identical70.5%57.4%Vals AIAs of 10/20/2026

Finance Agent (v2)Adjustments · higher is betterTest setup not documented as identical49.8%41.6%Vals AIAs of 10/29/2026

Legal Research BenchAdministrative / Regulatory · All-pass · higher is betterTest setup not documented as identical37.9%28.8%Vals AIAs of 11/05/2026

LMArena Code Arena WebDevHTML development · higher is betterTest setup not documented as identical1,5951,471LMArenaAs of 10/28/2026

LMArena Text, Style ControlBusiness, management and finance · higher is betterTest setup not documented as identical1,4411,448LMArenaAs of 10/28/2026

ProgramBenchAlmost Resolved · higher is betterTest setup not documented as identical8%0%Vals AIAs of 11/05/2026

Tax Agent BenchControversy & Precedence Analysis · higher is betterTest setup not documented as identical6.7%12.3%Vals AIAs of 11/05/2026

Terminal-Bench 4.0Hardware · higher is betterTest setup not documented as identical0%0%Vals AIAs of 11/05/2026

Code MigrationCOBOL · higher is betterTest setup not documented as identical58.6%58.4%Vals AIAs of 10/29/2026

EMBDataroom · higher is betterTest setup not documented as identical75.5%62.6%Vals AIAs of 10/20/2026

Finance Agent (v2)All-Pass · higher is betterTest setup not documented as identical38.4%31.4%Vals AIAs of 10/29/2026

Legal Research BenchAdministrative / Regulatory · Weighted partial credit (weighted pass) · higher is betterTest setup not documented as identical83.2%83.4%Vals AIAs of 11/05/2026

LMArena Code Arena WebDevReact development · higher is betterTest setup not documented as identical1,5751,454LMArenaAs of 10/28/2026

LMArena Text, Style ControlChinese · higher is betterTest setup not documented as identical1,4831,500LMArenaAs of 10/28/2026

ProgramBenchFully Resolved · higher is betterTest setup not documented as identical0.5%0.0%Vals AIAs of 11/05/2026

Tax Agent BenchControversy & Precedence Analysis (weighted pass) · higher is betterTest setup not documented as identical54.4%54.8%Vals AIAs of 11/05/2026

Terminal-Bench 4.0Media · higher is betterTest setup not documented as identical16.7%0.0%Vals AIAs of 11/05/2026

Code MigrationCode Quality · higher is betterTest setup not documented as identical3.752 / 53.401 / 5Vals AIAs of 10/29/2026

EMBDCF · higher is betterTest setup not documented as identical62.8%44.4%Vals AIAs of 10/20/2026

Finance Agent (v2)Comparables · higher is betterTest setup not documented as identical39.8%39.6%Vals AIAs of 10/29/2026

Legal Research BenchBusiness & Commercial · All-pass · higher is betterTest setup not documented as identical23.3%23.3%Vals AIAs of 11/05/2026

LMArena Code Arena WebDevWeb development · higher is betterTest setup not documented as identical1,5791,464LMArenaAs of 10/28/2026

LMArena Text, Style ControlCoding · higher is betterTest setup not documented as identical1,5061,488LMArenaAs of 10/28/2026

ProgramBenchRaw Pass Rate · higher is betterTest setup not documented as identical66.2%47.8%Vals AIAs of 11/05/2026

Tax Agent BenchCurrent & Temporal Analysis · higher is betterTest setup not documented as identical16.5%36.6%Vals AIAs of 11/05/2026

Terminal-Bench 4.0ML · higher is betterTest setup not documented as identical9.1%3.0%Vals AIAs of 11/05/2026

EMBLBO · higher is betterTest setup not documented as identical59.1%43.9%Vals AIAs of 10/20/2026

Finance Agent (v2)Disclosure Analysis · higher is betterTest setup not documented as identical57.6%52.6%Vals AIAs of 10/29/2026

Legal Research BenchBusiness & Commercial · Weighted partial credit (weighted pass) · higher is betterTest setup not documented as identical79.1%86.5%Vals AIAs of 11/05/2026

LMArena Text, Style ControlCreative writing · higher is betterTest setup not documented as identical1,4101,440LMArenaAs of 10/28/2026

Tax Agent BenchCurrent & Temporal Analysis (weighted pass) · higher is betterTest setup not documented as identical55.3%65.7%Vals AIAs of 11/05/2026

Terminal-Bench 4.0Operations · higher is betterTest setup not documented as identical3.7%7.4%Vals AIAs of 11/05/2026

EMBM&A · higher is betterTest setup not documented as identical72.1%48.2%Vals AIAs of 10/20/2026

Finance Agent (v2)Earnings Analysis · higher is betterTest setup not documented as identical74.2%59.8%Vals AIAs of 10/29/2026

Legal Research BenchCivil Litigation · All-pass · higher is betterTest setup not documented as identical25.9%33.3%Vals AIAs of 11/05/2026

LMArena Text, Style ControlEnglish · higher is betterTest setup not documented as identical1,4491,467LMArenaAs of 10/28/2026

More comparisons

Compare GPT-6 Luna and DeepSeek-V4-Pro with the models people compare them with most often.

Compare another matchup

First modelSecond model

Predecessor and successor

PredecessorGPT-6 Luna vs. GPT-5.6 LunaPredecessor of GPT-6 Luna

The most popular comparisons

Show all 68 comparisons