ZMIME
Comparison · 3 models · Updated Nov 22, 2026

DeepSeek V4.1 Flash vs Gemini 3.8 Flash vs GPT-5.6 Luna

Too close to call on our weighted score (DeepSeek V4.1 Flash 76, GPT-5.6 Luna 76, Gemini 3.8 Flash 70). The right pick depends on what you value most.Too close to call on our weighted score (Gemini 3.8 Flash 83, DeepSeek V4.1 Flash 81, GPT-5.6 Luna 81). The right pick depends on what you value most.

  1. DeepSeek

    DeepSeek V4.1 Flash

    Released Sep 10, 2026

    76/100
    • ECI157.0
    • Price$0.222 / $1.04
    • Context1M
  2. Google

    Gemini 3.8 Flash

    Released Sep 2, 2026Latest

    70/100
    • ECI158.8
    • Price$0.75 / $3.75
    • Context1.05M
  3. OpenAI

    GPT-5.6 Luna

    Released Jul 9, 2026

    76/100
    • ECI156.7
    • Price$0.20 / $1.20
    • Context1.05M
  4. DeepSeek

    DeepSeek V4.1 Flash

    Released Sep 10, 2026

    81/100
    • ECI157.0
    • Price$0.222 / $1.04
    • Context1M
  5. Google

    Gemini 3.8 Flash

    Released Sep 2, 2026Latest

    83/100
    • ECI158.8
    • Price$0.75 / $3.75
    • Context1.05M
  6. OpenAI

    GPT-5.6 Luna

    Released Jul 9, 2026

    81/100
    • ECI156.7
    • Price$0.20 / $1.20
    • Context1.05M
01 — Verdict

Too close to call

It is close. Our weighted score puts them within a point (DeepSeek V4.1 Flash 76/100, GPT-5.6 Luna 76/100, Gemini 3.8 Flash 70/100), so choose by what matters most for your work: Gemini 3.8 Flash for raw capability and DeepSeek V4.1 Flash on price. The score weighs capability 59%, price 29%, context window 12%.It is close. Our weighted score puts them within 2 points (Gemini 3.8 Flash 83/100, DeepSeek V4.1 Flash 81/100, GPT-5.6 Luna 81/100), so choose by what matters most for your work: Gemini 3.8 Flash for raw capability and DeepSeek V4.1 Flash on price. The score weighs capability 83%, context window 17%.

  • CapabilityGemini 3.8 FlashCapabilities Index (ECI): Gemini 3.8 Flash 158.8 · DeepSeek V4.1 Flash 157.0 · GPT-5.6 Luna 156.7
  • Lowest priceDeepSeek V4.1 FlashDeepSeek V4.1 Flash $0.426 · GPT-5.6 Luna $0.45 · Gemini 3.8 Flash $1.50 per 1M tokens (3:1 blend)
  • Longest contextGPT-5.6 Luna and Gemini 3.8 FlashGPT-5.6 Luna 1,050,000 · Gemini 3.8 Flash 1,048,576 · DeepSeek V4.1 Flash 1,000,000 tokens
  • Widest inputsGemini 3.8 FlashDeepSeek V4.1 Flash: Text, Images · Gemini 3.8 Flash: Text, Images, PDFs, Audio, Video · GPT-5.6 Luna: Text, Images, PDFs
  • Self-hostingDeepSeek V4.1 FlashPublishes downloadable weights (MIT)
How the score is built
MeasureWeightDeepSeek V4.1 FlashGemini 3.8 FlashGPT-5.6 Luna
CapabilityCapabilities Index (ECI)59%848684
Price29%654063
Context window12%646565
Overall100%76/10070/10076/100

Left out because at least one model lacks the data: inputs & features. The remaining weights were rescaled.

How the score is built
MeasureWeightDeepSeek V4.1 FlashGemini 3.8 FlashGPT-5.6 Luna
CapabilityCapabilities Index (ECI)83%848684
Context window17%646565
Overall100%81/10083/10081/100

Left out because at least one model lacks the data: inputs & features. The remaining weights were rescaled.

02 — Side by side

Every spec in one table

Highlighted cells lead their row. Unavailable specifications are marked “Not published”. Publisher results include their evaluation settings and source.

DeepSeek V4.1 Flash vs Gemini 3.8 Flash vs GPT-5.6 Luna specifications side by side
SpecificationDeepSeek V4.1 FlashDeepSeekGemini 3.8 FlashGoogleGPT-5.6 LunaOpenAI
Capability
Capabilities Index (ECI)157.0158.8 (best)156.7
ECI rank#35 of 159#25 of 159 (best)#37 of 159
ECI confidence interval155.1–159.2156.8–161.0154.9–158.8
ECI evaluation dateSep 9, 2026Sep 2, 2026Jul 9, 2026
Documented capabilitiesDeepSeek V4.1 Flash model for reasoning and agentic codingGoogle's most intelligent Flash model, engineered for long-horizon software engineering, autonomous agents, and complex enterprise workflowsCost-efficient GPT-5.6 model for fast, high-volume workloads
Model parametersNot publishedNot publishedNot published
Output modalitiesTextTextText
GPQA DiamondGraduate-level science questions89.8%max95.4%high (best)91.6%max
FrontierMath Tiers 1–3Research-level mathematics67.4%max68.4%high82.1%max (best)
OTIS Mock AIME 2024–2025Competition mathematics98.3%max98.9%high (best)98.3%max
SimpleQA VerifiedShort factual questions42.9%max69.7%high (best)41.0%max
Publisher-reported benchmarks
GPQA Diamondreasoning_effort=10090.9pass@1Reported resultNot publishedNot published
MathArena Apexreasoning_effort=10065.6pass@1Reported resultNot publishedNot published
Humanity's Last Exam · full setreasoning_effort=10036.8pass@1Reported resultNot publishedNot published
Humanity's Last Exam · text-only subsetreasoning_effort=10039.1pass@1Reported resultNot publishedNot published
Codeforcesreasoning_effort=1003471ratingReported resultNot publishedNot published
Terminal-Bench 2.1reasoning_effort=10090.6pass@1Reported resultNot publishedNot published
Terminal-Bench 3.0reasoning_effort=10030pass@1Reported resultNot publishedNot published
Terminal-Bench 4.0reasoning_effort=10031.2pass@1Reported resultNot publishedNot published
DeepSWE 1.1reasoning_effort=10074.2resolvedReported resultNot publishedNot published
ProgramBenchreasoning_effort=10020.3almost@1Reported resultNot publishedNot published
NL2Reporeasoning_effort=10064scoreReported resultNot publishedNot published
CyberGymreasoning_effort=10088.1pass@1Reported resultNot publishedNot published
ExploitGymreasoning_effort=10015.3pass@1Reported resultNot publishedNot published
AutomationBenchreasoning_effort=10054.8pass@1Reported resultNot publishedNot published
Agents' Last Examreasoning_effort=10031.8pass@1Reported resultNot publishedNot published
SEC-Bench Proreasoning_effort=10062.8pass@1Reported resultNot publishedNot published
Humanity's Last Examreasoning_effort=100; with tools63.9pass@1Reported resultNot publishedNot published
Chartographyreasoning_effort=100; with tools78.9pass@1Reported resultNot publishedNot published
BabyVisionreasoning_effort=100; with tools89.6pass@1Reported resultNot publishedNot published
ZeroBench · mainreasoning_effort=100; with tools49pass@5Reported resultNot publishedNot published
DeepSWE 1.1high thinkingNot published73.7pass@1Reported resultNot published
GDPval-AA 2Publisher-reported resultNot published1545EloReported resultNot published
Finance Agent 2Publisher-reported resultNot published61.4pass@1Reported resultNot published
Harvey Legal Agent BenchmarkPublisher-reported resultNot published10all-pass rateReported resultNot published
Terminal-Bench 2.1Publisher-reported resultNot published89.4pass@1Reported resultNot published
Terminal-Bench 4.0Publisher-reported resultNot published19.1pass@1Reported resultNot published
GDP.PDFPublisher-reported resultNot published35all-pass rateReported resultNot published
CharXiv Reasoningwithout toolsNot published86.2pass@1Reported resultNot published
LVBenchagenticNot published87.8pass@1Reported resultNot published
LVBenchstatic; 1024 frames; without toolsNot published87.1pass@1Reported resultNot published
HLE-Verified · 1811 verified and revised itemsPublisher-reported resultNot published54.9accuracyReported resultNot published
OSWorld 2.0 · before 2026-08-08 patchbatched tool calls; best of 3 runs; 500 stepsNot published59partial scoreReported resultNot published
BioMysteryBench · human-solvablewith toolsNot published88.8pass@1Reported resultNot published
BioMysteryBench · human-difficultwith toolsNot published56.5pass@1Reported resultNot published
LAB-Bench 2 · 11 subtaskswith toolsNot published86.2macro-averageReported resultNot published
SWE-Bench ProPublisher-reported resultNot publishedNot published62.7resolve rateReported resultJul 9, 2026
Terminal-Bench 2.1Publisher-reported resultNot publishedNot published84.7success rateReported resultJul 9, 2026
DeepSWE 1.1Publisher-reported resultNot publishedNot published67.2resolve rateReported resultJul 9, 2026
GPQA DiamondPublisher-reported resultNot publishedNot published92.3accuracyReported resultJul 9, 2026
FrontierMath v2 · Tier 1-3Publisher-reported resultNot publishedNot published78.6accuracyReported resultJul 9, 2026
BrowseCompPublisher-reported resultNot publishedNot published83.3accuracyReported resultJul 9, 2026
OSWorld 2.0Publisher-reported resultNot publishedNot published45.6success rateReported resultJul 9, 2026
MMMU Prono toolsNot publishedNot published78.4accuracyReported resultJul 9, 2026
Agents' Last ExamPublisher-reported resultNot publishedNot published50.3Reported resultJul 9, 2026
ToolathlonPublisher-reported resultNot publishedNot published53.4success rateReported resultJul 9, 2026
Artificial Analysis Intelligence Index 4.1maxNot publishedNot published51.2index scoreReported resultJul 9, 2026
Artificial Analysis Coding Agent Index 1.1maxNot publishedNot published74.6index scoreReported resultJul 9, 2026
Price per million tokens
Input$0.222$0.75$0.20 (best)
Output$1.04 (best)$3.75$1.20
Cached inputNot published$0.075$0.02 (best)
Blended (3:1)$0.426 (best)$1.50$0.45
Long-context rateSame rateSame rateOver 272K: $0.40 / $1.80
Price sourceMedian of 82 providersOfficial Google APIOfficial OpenAI API
Limits
Context window1,000,000 tokens1,048,576 tokens1,050,000 tokens (best)
Max inputNot publishedNot published922,000 tokens
Max output384,000 tokens (best)65,536 tokens128,000 tokens
Inputs and features
TextYesYesYes
ImagesYesYesYes
PDFsNoYesYes
AudioNoYesNo
VideoNoYesNo
ReasoningYesYeslow · medium · highYeslow · medium · high · xhigh · max
Tool callingYesYesYes
Structured outputNot publishedYesYes
File attachmentsYesYesYes
Temperature controlYesYesYes
Reasoning controlalwaysefforteffort
Availability
WeightsOpenMITProprietaryProprietary
API model IDNot publishedgemini-3.8-flashgpt-5.6-luna
API providers86 (best)4257
ReleasedSep 10, 2026Sep 2, 2026Jul 9, 2026
Knowledge cutoffMay 2025Not publishedFeb 16, 2026
Publisher documentationModel weightsPublisher documentationPublisher documentation
Source metadata updatedSep 10, 2026Sep 2, 2026Jul 9, 2026
Data sourcesmodels.dev · Epoch AImodels.dev · Epoch AImodels.dev · Epoch AI
Catalog checkedNov 22, 2026Nov 22, 2026Nov 22, 2026
03 — Cost

What would a month cost?

Enter your expected volume in millions of tokens. List prices only; caching and batch discounts would lower these.

  • DeepSeek V4.1 Flash$4.29
  • Gemini 3.8 Flash$15.00
  • GPT-5.6 Luna$4.40
04 — Questions

Which should you choose?

Which is better: DeepSeek V4.1 Flash, Gemini 3.8 Flash or GPT-5.6 Luna?

It is close. Our weighted score puts them within a point (DeepSeek V4.1 Flash 76/100, GPT-5.6 Luna 76/100, Gemini 3.8 Flash 70/100), so choose by what matters most for your work: Gemini 3.8 Flash for raw capability and DeepSeek V4.1 Flash on price. The score weighs capability 59%, price 29%, context window 12%.It is close. Our weighted score puts them within 2 points (Gemini 3.8 Flash 83/100, DeepSeek V4.1 Flash 81/100, GPT-5.6 Luna 81/100), so choose by what matters most for your work: Gemini 3.8 Flash for raw capability and DeepSeek V4.1 Flash on price. The score weighs capability 83%, context window 17%.

Which is cheaper, DeepSeek V4.1 Flash, Gemini 3.8 Flash or GPT-5.6 Luna?

DeepSeek V4.1 Flash is cheaper at $0.222 input / $1.04 output per million tokens (median across 82 API providers). GPT-5.6 Luna costs $0.20 input / $1.20 output per million tokens (official OpenAI API price); Gemini 3.8 Flash costs $0.75 input / $3.75 output per million tokens (official Google API price). At a typical mix of three input tokens to one output token, that is $0.426 per million tokens for DeepSeek V4.1 Flash versus $0.45 for GPT-5.6 Luna (1.1× as much) and $1.50 for Gemini 3.8 Flash (3.5× as much).

Which scores higher on benchmarks?

Gemini 3.8 Flash scores higher on the Capabilities Index (ECI): Gemini 3.8 Flash 158.8 (#25 of 159), DeepSeek V4.1 Flash 157.0 (#35 of 159) and GPT-5.6 Luna 156.7 (#37 of 159). The confidence ranges of the top two overlap (156.8–161.0 vs 155.1–159.2), so treat the gap as small. On individual benchmarks: GPQA Diamond — Gemini 3.8 Flash 95.4%, GPT-5.6 Luna 91.6%, DeepSeek V4.1 Flash 89.8%; FrontierMath Tiers 1–3 — GPT-5.6 Luna 82.1%, Gemini 3.8 Flash 68.4%, DeepSeek V4.1 Flash 67.4%; OTIS Mock AIME 2024–2025 — Gemini 3.8 Flash 98.9%, DeepSeek V4.1 Flash 98.3%, GPT-5.6 Luna 98.3%; SimpleQA Verified — Gemini 3.8 Flash 69.7%, DeepSeek V4.1 Flash 42.9%, GPT-5.6 Luna 41.0%.

Which is better for coding?

There are no published SWE-bench Verified results for DeepSeek V4.1 Flash, Gemini 3.8 Flash and GPT-5.6 Luna yet, so there is no like-for-like coding score. On overall capability, Gemini 3.8 Flash leads, which tends to carry over to coding, but test on your own codebase. All three support tool calling for agent workflows.

Which has the bigger context window?

GPT-5.6 Luna and Gemini 3.8 Flash have the largest context windows (1,050,000 and 1,048,576 tokens), against 1,000,000 for DeepSeek V4.1 Flash. Maximum output per response: DeepSeek V4.1 Flash up to 384,000, Gemini 3.8 Flash up to 65,536, GPT-5.6 Luna up to 128,000 tokens.

Which can read images, PDFs, audio or video?

DeepSeek V4.1 Flash accepts text and images; Gemini 3.8 Flash accepts text, images, PDFs, audio and video; GPT-5.6 Luna accepts text, images and PDFs. Gemini 3.8 Flash handles the widest range of inputs.

Are any of these open source?

DeepSeek V4.1 Flash publishes its weights (MIT) and can be self-hosted; Gemini 3.8 Flash and GPT-5.6 Luna is proprietary.

Which is newer?

DeepSeek V4.1 Flash is the newest, released Sep 10, 2026. Gemini 3.8 Flash came out Sep 2, 2026; GPT-5.6 Luna came out Jul 9, 2026. Knowledge cutoff: DeepSeek V4.1 Flash May 2025, GPT-5.6 Luna Feb 16, 2026.

How do you decide the winner?

This rating uses capability (59%, independent benchmark results); price (29%, blended price per million tokens (3 input : 1 output), log scale); context window (12%, maximum tokens per request, log scale). Dimensions missing for any model are dropped and the remaining weights rescaled, so every model is judged on the same evidence. Specs and prices come from public model listings and the labs’ own API pages; capability scores come from independent benchmark runs. Data updated Nov 22, 2026.This rating uses capability (83%, independent benchmark results); context window (17%, maximum tokens per request, log scale). Dimensions missing for any model are dropped and the remaining weights rescaled, so every model is judged on the same evidence. Specs and prices come from public model listings and the labs’ own API pages; capability scores come from independent benchmark runs. Data updated Nov 22, 2026.