Skip to main content
BenchLM

DeepSeek V3 vs GPT-5.6 Luna

Updated September 24, 2026. Rank says GPT-5.6 Luna is ahead. Price, access, and your workload can each overturn that. Public scores include evidence status and uncertainty.

Share or export
Share on XLinkedInSocial cardCSVJSON

Decision reading

GPT-5.6 Luna has the higher public score, 65.6 versus 31.9, and the 90% score intervals do not overlap. 2 results are shared. Category rows resting on Estimated evidence or different benchmark sets are marked directional and do not name a winner.

Model A
DeepSeek logo
DeepSeek V3

DeepSeek

31.9/100

Supported · Public rank #145

90% interval 15.9–47.9

Model B
OpenAI logo
GPT-5.6 Luna

OpenAI

65.6/100

Supported · Public rank #25

90% interval 60.5–70.7

Shared results
2
DeepSeek V3 only
4
GPT-5.6 Luna only
27
Like-for-like categories
0 / 8
Supported: DeepSeek V3 and GPT-5.6 LunaHow the comparison works

Which one for your work

Recommendations appear only when a shared evidence basis or an explicit operating constraint supports the call. Secondary and unsupported use cases stay disclosed below the initial list.

  • Long documents

    Prompts that approach the documented context limit

    GPT-5.6 Luna

    GPT-5.6 Luna has the larger documented context window.

    Confidence: documented

  • Chat turn cost

    1K fresh input + 500 output tokens

    GPT-5.6 Luna

    GPT-5.6 Luna has the lower estimated token cost for this stated workload. Costs use the listed standard API rates.

    Confidence: listed-rates
Show secondary and unsupported calls

  • Repository review cost

    50K fresh input + 3K output tokens

    GPT-5.6 Luna

    GPT-5.6 Luna has the lower estimated token cost for this stated workload. Costs use the listed standard API rates.

    Confidence: listed-rates

  • Coding work

    Code generation, repair, and software-engineering tasks

    Not enough matched evidence

    DeepSeek V3 is scored on Estimated evidence for coding, so the reading is directional rather than like-for-like.

    Confidence: limited

  • Agentic work

    Tool use, computer use, and multi-step task completion

    Not enough matched evidence

    DeepSeek V3 is scored on Estimated evidence for agentic, so the reading is directional rather than like-for-like.

    Confidence: limited

  • Cache-heavy agent loop cost

    200K cached + 20K fresh input + 10K output tokens

    Not enough matched evidence

    The page does not recommend a cost winner because at least one model cannot fit the stated workload in one request. DeepSeek V3 does not fit this workload in one request.

    Confidence: listed-rates

Which one for a specific job

Choose a job from the LLM Selector's task catalog to see the category row it rests on, under the same basis rules as the table below. A directional row stays directional; choosing a job never creates a winner.

The same task catalog as the LLM Selector. Each job names the evidence surface it rests on; nothing here adds a new score.

18.9DeepSeek V364.5GPT-5.6 Luna

Directional only · BenchAlign v5.7

GPT-5.6 Luna scores higher, but at least one score rests on Estimated evidence or a different benchmark set. Directional only, no winner.

Coding scores combine specific tasks and setups. Match the editor, harness, and effort to your workflow.

Open the coding leaderboardShortlist models for this job
Same basis rules as the category table below

What is actually comparable

Shared results can support a head-to-head reading. Results present for only one model describe coverage, not superiority.

3 categories rest on Estimated evidence or different evidence sets. Those rows remain visible for coverage context but do not name a winner.

Shape of the matched evidence

Only shared public evidence is shown. Sparse evidence stays a ruled list rather than being closed into a radar shape.

Too few matched category axes support a radar. The ruled list below shows only shared benchmark results; positions use each benchmark’s normalized display scale when available.

Bars run 0–100 on each benchmark’s normalized display scale

Category results, on a stated basis

Each row shows the public-lane category score for both models: the BenchAlign v5.7 lane for agentic, coding, and knowledge, and the weighted public rows elsewhere, with the same rank each model page reports. A row is like-for-like only when both scores rest on Supported evidence or the same weighted set. Directional and not-comparable rows remain visible, but they never receive a winner.

Agentic

Directional only
DeepSeek V3
11.1
Estimated · #102/105
GPT-5.6 Luna
55.3
Supported · #28/105
Basis
BenchAlign v5.7 lane · 0 vs 9 public rows
Reading
Directional only

Coding

Directional only
DeepSeek V3
18.9
Estimated · #123/135
GPT-5.6 Luna
64.5
Supported · #9/135
Basis
BenchAlign v5.7 lane · 2 vs 7 public rows
Reading
Directional only

Knowledge

Directional only
DeepSeek V3
30.0
Estimated · #130/158
GPT-5.6 Luna
64.6
Supported · #22/158
Basis
BenchAlign v5.7 lane · 2 vs 6 public rows
Reading
Directional only

Reasoning

Not comparable
DeepSeek V3
42.2
Unranked · 2 rankable rows
GPT-5.6 Luna
54.7
#18/19
Basis
Provisional lane · 0 vs 2 weighted rows
Reading
Not comparable

Multimodal

Not comparable
DeepSeek V3
Not ranked
GPT-5.6 Luna
67.1
#22/50
Basis
Provisional lane · 0 vs 1 weighted rows
Reading
Not comparable

Multilingual

Not comparable
DeepSeek V3
Not ranked
GPT-5.6 Luna
Not ranked
Basis
Provisional lane · 0 vs 0 weighted rows
Reading
Not comparable

Instruction following

Not comparable
DeepSeek V3
38.2
#103/124
GPT-5.6 Luna
Not ranked
Basis
Provisional lane · 0 vs 0 weighted rows
Reading
Not comparable

Math

Not comparable
DeepSeek V3
26.0
Unranked · 1 rankable row
GPT-5.6 Luna
94.2
Unranked · 3 rankable rows
Basis
Provisional lane · 1 vs 2 weighted rows
Reading
Not comparable

Ranks count the models scored in each category’s lane, so the agentic, coding, and knowledge denominators (BenchAlign v5.7) differ from the provisional-lane categories. Unranked scores sit on the provisional lane’s scale but never name a winner. Like-for-like rows within 0.5 points read as a practical tie.

Supported evidence per lane · bars run 0–100Methodology

What each workload costs

Three fixed token mixes turn per-token rates into comparable decisions. Each scenario states context fit and whether cached input had to fall back to the published list-input rate.

Chat turn

1K fresh input + 500 output tokens

DeepSeek V3
$0.00082
Fits in one request
GPT-5.6 Luna
$0.0008
Fits in one request

GPT-5.6 Luna has the lower modeled cost

Costs use the listed standard API rates.

Repository review

50K fresh input + 3K output tokens

DeepSeek V3
$0.0168
Fits in one request
GPT-5.6 Luna
$0.0136
Fits in one request

GPT-5.6 Luna has the lower modeled cost

Costs use the listed standard API rates.

Cache-heavy agent loop

200K cached + 20K fresh input + 10K output tokens

DeepSeek V3
$0.0304
Does not fit in one request
GPT-5.6 Luna
$0.02
Fits in one request

DeepSeek V3 does not fit this workload in one request.

Cached input falls back to the list input rate only where a cached rate is unpublished

Specification differences

Sourced differences are shown directly. Missing facts stay explicit instead of being inferred from a model name or family.

Context window

Maximum documented context; output-token limits may be lower.

DeepSeek V3

128K

GPT-5.6 Luna

1.05M

OpenAI model catalog

Cached-input rate

A missing cached-input rate falls back to the listed input rate only in the stated workload estimate.

DeepSeek V3

$0.07 per 1M cached input tokens

GPT-5.6 Luna

$0.02 per 1M cached input tokens

OpenAI pricing

Provider availability

DeepSeek V3

Not sourced

GPT-5.6 Luna

Generally Available · OpenAI Responses API

OpenAI model catalog

Reasoning profile

DeepSeek V3

Non-Reasoning

GPT-5.6 Luna

Reasoning

Weight access

DeepSeek V3

Open Weight

GPT-5.6 Luna

Proprietary

License

DeepSeek V3

Open Weight

GPT-5.6 Luna

Proprietary

Release date

DeepSeek V3

2024-12-26

GPT-5.6 Luna

2026-07-09

If you already use one of these models

Deployment change
The models list different providers, so authentication, endpoint behavior, limits, and feature support may change.
Quality signal
GPT-5.6 Luna has the higher public score, 65.6 versus 31.9, and the 90% score intervals do not overlap.
Workload cost
Repository review: $0.0168 vs $0.0136. Cache-heavy agent loop: $0.0304 vs $0.02.
Context tradeoff
GPT-5.6 Luna has the larger documented window (1.05M).
Run the same representative tasks against both endpoints before changing production traffic.

Questions

Which is better, DeepSeek V3 or GPT-5.6 Luna?

GPT-5.6 Luna has the higher public score, 65.6 versus 31.9, and the 90% score intervals do not overlap. That is the clearest overall quality signal in the current public data.

Which is better for coding, DeepSeek V3 or GPT-5.6 Luna?

GPT-5.6 Luna scores higher for coding on the public lane, 64.5 to 18.9. DeepSeek V3 is scored on Estimated evidence for coding, so the reading is directional rather than like-for-like. BenchLM does not name a winner for coding from a directional reading; read the shared benchmark rows directly and test the models on the same task set.

Which is better for agentic tasks, DeepSeek V3 or GPT-5.6 Luna?

GPT-5.6 Luna scores higher for agentic tasks on the public lane, 55.3 to 11.1. DeepSeek V3 is scored on Estimated evidence for agentic tasks, so the reading is directional rather than like-for-like. BenchLM does not name a winner for agentic tasks from a directional reading; read the shared benchmark rows directly and test the models on the same task set.

Which costs less, DeepSeek V3 or GPT-5.6 Luna?

For the stated presets, chat costs $0.00082 on DeepSeek V3 and $0.0008 on GPT-5.6 Luna; repository review costs $0.0168 and $0.0136; the cache-heavy agent loop costs $0.0304 and $0.02. DeepSeek V3 does not fit this workload in one request.

Which has the larger context window, DeepSeek V3 or GPT-5.6 Luna?

GPT-5.6 Luna has the larger documented context window: 1.05M, compared with 128K.

Self-host vs API cost

Estimates at 50,000 req/day · 1000 tokens/req average.

DeepSeek V3
API / mo$1,028
Self-host / mo$18,221
Break-even1.2B/day
GPT-5.6 Luna
API / mo$1,050
Self-host / moNot listed
Break-even—
Proprietary model — self-hosting not applicable.
Model the full break-even

Benchmark evidence

The full public result ledger is available for audit without forcing a wide desktop table onto a phone.

Browse raw public benchmark evidence33 rows

Agentic

  • Terminal-Bench 3.0

    DeepSeek V3—
    GPT-5.6 Luna14.3%
    Source

    Not directly comparable

  • Terminal-Bench 2.1

    DeepSeek V3—
    GPT-5.6 Luna84.7%
    Source

    Not directly comparable

  • BrowseComp

    DeepSeek V3—
    GPT-5.6 Luna83.3%
    Source

    Not directly comparable

  • OSWorld 2.0

    DeepSeek V3—
    GPT-5.6 Luna45.6%
    Source

    Not directly comparable

  • CyberGym

    DeepSeek V3—
    GPT-5.6 Luna77.9%
    Source

    Not directly comparable

  • ExploitGym

    DeepSeek V3—
    GPT-5.6 Luna12.4%
    Source

    Not directly comparable

  • Toolathlon

    DeepSeek V3—
    GPT-5.6 Luna53.4%
    Source

    Not directly comparable

  • Terminal-Bench 2.1 (Vals)

    DeepSeek V3—
    GPT-5.6 Luna79.0%
    Source

    Not directly comparable

  • ApprenticeBench

    DeepSeek V3—
    GPT-5.6 Luna7%
    Source

    Not directly comparable

Coding

  • LiveCodeBench

    DeepSeek V337.6%
    Source
    GPT-5.6 Luna—

    Not directly comparable

  • SWE-bench Verified

    DeepSeek V342%
    Source
    GPT-5.6 Luna—

    Not directly comparable

  • SWE-bench Pro

    DeepSeek V3—
    GPT-5.6 Luna62.7%
    Source

    Not directly comparable

  • Terminal-Bench 2.1

    DeepSeek V3—
    GPT-5.6 Luna84.7%
    Source

    Not directly comparable

  • DeepSWE

    DeepSeek V3—
    GPT-5.6 Luna67.2%
    Source

    Not directly comparable

  • FrontierCode 1.1 Extended

    DeepSeek V3—
    GPT-5.6 Luna55.1%
    Source

    Not directly comparable

  • cursorBench32

    DeepSeek V3—
    GPT-5.6 Luna61.1%
    Source

    Not directly comparable

  • VulcanBench v3

    DeepSeek V3—
    GPT-5.6 Luna85.5%
    Source

    Not directly comparable

  • SWE-bench (Vals)

    DeepSeek V3—
    GPT-5.6 Luna93.0%
    Source

    Not directly comparable

Reasoning

  • ARC-AGI-2

    DeepSeek V3—
    GPT-5.6 Luna59.5%
    Source

    Not directly comparable

  • ARC-AGI-3

    DeepSeek V3—
    GPT-5.6 Luna0.2%
    Source

    Not directly comparable

Multimodal

  • MMMU-Pro

    DeepSeek V3—
    GPT-5.6 Luna78.4%
    Source

    Not directly comparable

  • MMMU-Pro w/ Python

    DeepSeek V3—
    GPT-5.6 Luna79.5%
    Source

    Not directly comparable

Knowledge

  • GPQA

    DeepSeek V359.1%
    Source
    GPT-5.6 Luna92.3%
    Source

    GPT-5.6 Luna leads this result

  • MMLU-Pro

    DeepSeek V375.9%
    Source
    GPT-5.6 Luna—

    Not directly comparable

  • GPQA-D

    DeepSeek V3—
    GPT-5.6 Luna92.3%
    Source

    Not directly comparable

  • HealthBench Professional

    DeepSeek V3—
    GPT-5.6 Luna55.7%
    Source

    Not directly comparable

  • HealthBench Hard

    DeepSeek V3—
    GPT-5.6 Luna32.0%
    Source

    Not directly comparable

  • GPQA Diamond (Vals)

    DeepSeek V3—
    GPT-5.6 Luna91.7%
    Source

    Not directly comparable

  • MMLU-Pro (Vals)

    DeepSeek V3—
    GPT-5.6 Luna86.0%
    Source

    Not directly comparable

Instruction following

  • IFEval

    DeepSeek V386.1%
    Source
    GPT-5.6 Luna—

    Not directly comparable

Math

  • FrontierMath v2 (Tiers 1-3)

    DeepSeek V31.724%
    Source
    GPT-5.6 Luna78.600%
    Source

    GPT-5.6 Luna leads this result

  • FrontierMath (legacy)

    DeepSeek V3—
    GPT-5.6 Luna78.6%
    Source

    Not directly comparable

  • FrontierMath v2 (Tier 4)

    DeepSeek V3—
    GPT-5.6 Luna58.500%
    Source

    Not directly comparable

33 public results · 2 shared

Related comparisons

Compare API pricingRead the methodologyOpen the model selector

Watch DeepSeek V3 vs GPT-5.6 Luna

One weekly email when material rank, price, or benchmark evidence changes make this matchup worth revisiting.

Read a sample issue

Join 2,000+ readers.

Last updated September 24, 2026