DeepSeek, released Sep 10, 2026

DeepSeek V4.1 Flashprice, context, benchmarks and release details

Open weights 69% confidence 69 percent, Medium confidence, 4 of 7 expected sources in
66.2
SI Score
#31 of 142 ranked models
Input, per 1M tokens
$0.15DeepSeek pricingOfficial off-peak uncached rate; peak is 2x; time schedule at source Retrieved Oct 9, 2026 · factual citation
Open source ↗
Output, per 1M tokens
$0.60DeepSeek pricingOfficial off-peak uncached rate; peak is 2x; time schedule at source Retrieved Oct 9, 2026 · factual citation
Open source ↗
Context window
1Mmodels.devPublished source fact Retrieved Oct 9, 2026 · MIT
Open source ↗
Max output
384Kmodels.devPublished source fact Retrieved Oct 9, 2026 · MIT
Open source ↗
Released
Sep 10, 2026DeepSeek V4.1 Flash announcementOfficial dated introduction and availability announcement Retrieved Oct 9, 2026 · factual citation
Open source ↗

How the score breaks down

Coding (weight 40 percent) 73.3
Math (weight 15 percent) 88.0
Preference (weight 15 percent) 78.8
Reasoning (weight 30 percent) 82.7

Weights: reasoning 30%, math 15%, coding 40%, preference 15%. Results use fixed 0–100 scales before averaging, and thin evidence is pulled toward 50. Method si-v3-retained-evidence-2, computed Oct 9, 2026, 05:13 UTC.

Around it on the leaderboard

  1. 29 Qwen3.6 Max Preview 66.4
  2. 30 GLM-5.3-Flash 66.2
  3. 31 DeepSeek V4.1 Flash 66.2
  4. 32 Kimi K2 Thinking Turbo 66.1
  5. 33 GPT-5.2 65.9

Full leaderboard

Benchmark results

28 benchmarks, 43 results

Each row shows the best published result. Where a model was tested at several settings, such as reasoning effort, open the row to see each one. Hover or tap a value for its source.

ARC-AGI-1 (public eval)reasoning 98%ARC PrizeARC Prize steward-published result; exact edition/subset; publication time is HTTP Last-Modified of the aggregate export, not the evaluation date. Evaluation conducted with 'max' reasoning effort using the standard harness. Partial dataset coverage: scores use full catalog denominators; costs include recorded usage only, excluding unrecorded attempts. [variant] deepseek-v4-1-flash-maxPublished Oct 6, 2026 Retrieved Oct 9, 2026 · factual citation
Open source ↗ 98.0 3 settings
  • low effort 96.3%ARC PrizeARC Prize steward-published result; exact edition/subset; publication time is HTTP Last-Modified of the aggregate export, not the evaluation date. Evaluation conducted with 'low' reasoning effort using the standard harness. [variant] deepseek-v4-1-flash-lowPublished Oct 6, 2026 Retrieved Oct 9, 2026 · factual citation
    Open source ↗
  • high effort 95.9%ARC PrizeARC Prize steward-published result; exact edition/subset; publication time is HTTP Last-Modified of the aggregate export, not the evaluation date. Evaluation conducted with 'high' reasoning effort using the standard harness. Partial dataset coverage: scores use full catalog denominators; costs include recorded usage only, excluding unrecorded attempts. [variant] deepseek-v4-1-flash-highPublished Oct 6, 2026 Retrieved Oct 9, 2026 · factual citation
    Open source ↗
  • max effort 98%ARC PrizeARC Prize steward-published result; exact edition/subset; publication time is HTTP Last-Modified of the aggregate export, not the evaluation date. Evaluation conducted with 'max' reasoning effort using the standard harness. Partial dataset coverage: scores use full catalog denominators; costs include recorded usage only, excluding unrecorded attempts. [variant] deepseek-v4-1-flash-maxPublished Oct 6, 2026 Retrieved Oct 9, 2026 · factual citation
    Open source ↗
About ARC-AGI-1 (public eval)
ARC-AGI-1 (semi-private)reasoning 94.5%ARC PrizeARC Prize steward-published result; exact edition/subset; publication time is HTTP Last-Modified of the aggregate export, not the evaluation date. Evaluation conducted with 'max' reasoning effort using the standard harness. Partial dataset coverage: scores use full catalog denominators; costs include recorded usage only, excluding unrecorded attempts. [variant] deepseek-v4-1-flash-maxPublished Oct 6, 2026 Retrieved Oct 9, 2026 · factual citation
Open source ↗ 94.5 3 settings
  • low effort 90.5%ARC PrizeARC Prize steward-published result; exact edition/subset; publication time is HTTP Last-Modified of the aggregate export, not the evaluation date. Evaluation conducted with 'low' reasoning effort using the standard harness. [variant] deepseek-v4-1-flash-lowPublished Oct 6, 2026 Retrieved Oct 9, 2026 · factual citation
    Open source ↗
  • high effort 88.5%ARC PrizeARC Prize steward-published result; exact edition/subset; publication time is HTTP Last-Modified of the aggregate export, not the evaluation date. Evaluation conducted with 'high' reasoning effort using the standard harness. Partial dataset coverage: scores use full catalog denominators; costs include recorded usage only, excluding unrecorded attempts. [variant] deepseek-v4-1-flash-highPublished Oct 6, 2026 Retrieved Oct 9, 2026 · factual citation
    Open source ↗
  • max effort 94.5%ARC PrizeARC Prize steward-published result; exact edition/subset; publication time is HTTP Last-Modified of the aggregate export, not the evaluation date. Evaluation conducted with 'max' reasoning effort using the standard harness. Partial dataset coverage: scores use full catalog denominators; costs include recorded usage only, excluding unrecorded attempts. [variant] deepseek-v4-1-flash-maxPublished Oct 6, 2026 Retrieved Oct 9, 2026 · factual citation
    Open source ↗
About ARC-AGI-1 (semi-private)
ARC-AGI-2 (public eval)reasoning 81.7%ARC PrizeARC Prize steward-published result; exact edition/subset; publication time is HTTP Last-Modified of the aggregate export, not the evaluation date. Evaluation conducted with 'max' reasoning effort using the standard harness. Partial dataset coverage: scores use full catalog denominators; costs include recorded usage only, excluding unrecorded attempts. [variant] deepseek-v4-1-flash-maxPublished Oct 6, 2026 Retrieved Oct 9, 2026 · factual citation
Open source ↗ 81.7 3 settings
  • low effort 66.8%ARC PrizeARC Prize steward-published result; exact edition/subset; publication time is HTTP Last-Modified of the aggregate export, not the evaluation date. Evaluation conducted with 'low' reasoning effort using the standard harness. [variant] deepseek-v4-1-flash-lowPublished Oct 6, 2026 Retrieved Oct 9, 2026 · factual citation
    Open source ↗
  • high effort 71.9%ARC PrizeARC Prize steward-published result; exact edition/subset; publication time is HTTP Last-Modified of the aggregate export, not the evaluation date. Evaluation conducted with 'high' reasoning effort using the standard harness. Partial dataset coverage: scores use full catalog denominators; costs include recorded usage only, excluding unrecorded attempts. [variant] deepseek-v4-1-flash-highPublished Oct 6, 2026 Retrieved Oct 9, 2026 · factual citation
    Open source ↗
  • max effort 81.7%ARC PrizeARC Prize steward-published result; exact edition/subset; publication time is HTTP Last-Modified of the aggregate export, not the evaluation date. Evaluation conducted with 'max' reasoning effort using the standard harness. Partial dataset coverage: scores use full catalog denominators; costs include recorded usage only, excluding unrecorded attempts. [variant] deepseek-v4-1-flash-maxPublished Oct 6, 2026 Retrieved Oct 9, 2026 · factual citation
    Open source ↗
About ARC-AGI-2 (public eval)
ARC-AGI-2 (semi-private)reasoning 72.9%ARC PrizeARC Prize steward-published result; exact edition/subset; publication time is HTTP Last-Modified of the aggregate export, not the evaluation date. Evaluation conducted with 'max' reasoning effort using the standard harness. Partial dataset coverage: scores use full catalog denominators; costs include recorded usage only, excluding unrecorded attempts. [variant] deepseek-v4-1-flash-maxPublished Oct 6, 2026 Retrieved Oct 9, 2026 · factual citation
Open source ↗ 72.9 3 settings
  • low effort 60.6%ARC PrizeARC Prize steward-published result; exact edition/subset; publication time is HTTP Last-Modified of the aggregate export, not the evaluation date. Evaluation conducted with 'low' reasoning effort using the standard harness. [variant] deepseek-v4-1-flash-lowPublished Oct 6, 2026 Retrieved Oct 9, 2026 · factual citation
    Open source ↗
  • high effort 67.5%ARC PrizeARC Prize steward-published result; exact edition/subset; publication time is HTTP Last-Modified of the aggregate export, not the evaluation date. Evaluation conducted with 'high' reasoning effort using the standard harness. Partial dataset coverage: scores use full catalog denominators; costs include recorded usage only, excluding unrecorded attempts. [variant] deepseek-v4-1-flash-highPublished Oct 6, 2026 Retrieved Oct 9, 2026 · factual citation
    Open source ↗
  • max effort 72.9%ARC PrizeARC Prize steward-published result; exact edition/subset; publication time is HTTP Last-Modified of the aggregate export, not the evaluation date. Evaluation conducted with 'max' reasoning effort using the standard harness. Partial dataset coverage: scores use full catalog denominators; costs include recorded usage only, excluding unrecorded attempts. [variant] deepseek-v4-1-flash-maxPublished Oct 6, 2026 Retrieved Oct 9, 2026 · factual citation
    Open source ↗
About ARC-AGI-2 (semi-private)
GPQA Diamondreasoning 90.9%Official model cards via models.devLab-reported; metric pass@1; transcribed by MIT models.dev catalog; not independently evaluated [variant] reasoning_effort=100 Retrieved Oct 9, 2026 · factual citation; MIT transcription
Open source ↗ 90.9
Humanity's Last Exam (full set)reasoning 36.8%Official model cards via models.devLab-reported; metric pass@1; transcribed by MIT models.dev catalog; not independently evaluated [variant] reasoning_effort=100; full set Retrieved Oct 9, 2026 · factual citation; MIT transcription
Open source ↗ 36.8
Humanity's Last Exam (text-only subset)reasoning 39.1%Official model cards via models.devLab-reported; metric pass@1; transcribed by MIT models.dev catalog; not independently evaluated [variant] reasoning_effort=100; text-only subset Retrieved Oct 9, 2026 · factual citation; MIT transcription
Open source ↗ 39.1
Humanity's Last Exam (with tools)reasoning 63.9%Official model cards via models.devLab-reported; metric pass@1; transcribed by MIT models.dev catalog; not independently evaluated [variant] reasoning_effort=100; with tools Retrieved Oct 9, 2026 · factual citation; MIT transcription
Open source ↗ 63.9
LiveBench Reasoning: connectionsreasoning 100%LiveBenchLiveBench subtask result; dataset edition 2026_06_25; publication time is HTTP Last-Modified of the score CSV, not the dataset edition or individual evaluation date [variant] 2026_06_25Published Oct 7, 2026 Retrieved Oct 9, 2026 · factual citation; Apache-2.0 code
Open source ↗ 100.0
LiveBench Reasoning: consecutive eventsreasoning 88.7%LiveBenchLiveBench subtask result; dataset edition 2026_06_25; publication time is HTTP Last-Modified of the score CSV, not the dataset edition or individual evaluation date [variant] 2026_06_25Published Oct 7, 2026 Retrieved Oct 9, 2026 · factual citation; Apache-2.0 code
Open source ↗ 88.7
LiveBench Reasoning: logic with navigationreasoning 68%LiveBenchLiveBench subtask result; dataset edition 2026_06_25; publication time is HTTP Last-Modified of the score CSV, not the dataset edition or individual evaluation date [variant] 2026_06_25Published Oct 7, 2026 Retrieved Oct 9, 2026 · factual citation; Apache-2.0 code
Open source ↗ 68.0
LiveBench Reasoning: spatialreasoning 98%LiveBenchLiveBench subtask result; dataset edition 2026_06_25; publication time is HTTP Last-Modified of the score CSV, not the dataset edition or individual evaluation date [variant] 2026_06_25Published Oct 7, 2026 Retrieved Oct 9, 2026 · factual citation; Apache-2.0 code
Open source ↗ 98.0
LiveBench Reasoning: theory of mindreasoning 80.8%LiveBenchLiveBench subtask result; dataset edition 2026_06_25; publication time is HTTP Last-Modified of the score CSV, not the dataset edition or individual evaluation date [variant] 2026_06_25Published Oct 7, 2026 Retrieved Oct 9, 2026 · factual citation; Apache-2.0 code
Open source ↗ 80.8
LiveBench Reasoning: zebra puzzlesreasoning 100%LiveBenchLiveBench subtask result; dataset edition 2026_06_25; publication time is HTTP Last-Modified of the score CSV, not the dataset edition or individual evaluation date [variant] 2026_06_25Published Oct 7, 2026 Retrieved Oct 9, 2026 · factual citation; Apache-2.0 code
Open source ↗ 100.0
LiveBench Math: AMPS Hardmath 98%LiveBenchLiveBench subtask result; dataset edition 2026_06_25; publication time is HTTP Last-Modified of the score CSV, not the dataset edition or individual evaluation date [variant] 2026_06_25Published Oct 7, 2026 Retrieved Oct 9, 2026 · factual citation; Apache-2.0 code
Open source ↗ 98.0
LiveBench Math: competition mathmath 96.1%LiveBenchLiveBench subtask result; dataset edition 2026_06_25; publication time is HTTP Last-Modified of the score CSV, not the dataset edition or individual evaluation date [variant] 2026_06_25Published Oct 7, 2026 Retrieved Oct 9, 2026 · factual citation; Apache-2.0 code
Open source ↗ 96.1
LiveBench Math: integralsmath 90%LiveBenchLiveBench subtask result; dataset edition 2026_06_25; publication time is HTTP Last-Modified of the score CSV, not the dataset edition or individual evaluation date [variant] 2026_06_25Published Oct 7, 2026 Retrieved Oct 9, 2026 · factual citation; Apache-2.0 code
Open source ↗ 90.0
LiveBench Math: olympiadmath 89.1%LiveBenchLiveBench subtask result; dataset edition 2026_06_25; publication time is HTTP Last-Modified of the score CSV, not the dataset edition or individual evaluation date [variant] 2026_06_25Published Oct 7, 2026 Retrieved Oct 9, 2026 · factual citation; Apache-2.0 code
Open source ↗ 89.1
LiveBench Math: simplifymath 67.1%LiveBenchLiveBench subtask result; dataset edition 2026_06_25; publication time is HTTP Last-Modified of the score CSV, not the dataset edition or individual evaluation date [variant] 2026_06_25Published Oct 7, 2026 Retrieved Oct 9, 2026 · factual citation; Apache-2.0 code
Open source ↗ 67.1
LiveBench Coding: code completioncoding 82.6%LiveBenchLiveBench subtask result; dataset edition 2026_06_25; publication time is HTTP Last-Modified of the score CSV, not the dataset edition or individual evaluation date [variant] 2026_06_25Published Oct 7, 2026 Retrieved Oct 9, 2026 · factual citation; Apache-2.0 code
Open source ↗ 82.6
LiveBench Coding: code generationcoding 77.5%LiveBenchLiveBench subtask result; dataset edition 2026_06_25; publication time is HTTP Last-Modified of the score CSV, not the dataset edition or individual evaluation date [variant] 2026_06_25Published Oct 7, 2026 Retrieved Oct 9, 2026 · factual citation; Apache-2.0 code
Open source ↗ 77.5
LiveBench Coding: JavaScriptcoding 81.8%LiveBenchLiveBench subtask result; dataset edition 2026_06_25; publication time is HTTP Last-Modified of the score CSV, not the dataset edition or individual evaluation date [variant] 2026_06_25Published Oct 7, 2026 Retrieved Oct 9, 2026 · factual citation; Apache-2.0 code
Open source ↗ 81.8
LiveBench Coding: Pythoncoding 90%LiveBenchLiveBench subtask result; dataset edition 2026_06_25; publication time is HTTP Last-Modified of the score CSV, not the dataset edition or individual evaluation date [variant] 2026_06_25Published Oct 7, 2026 Retrieved Oct 9, 2026 · factual citation; Apache-2.0 code
Open source ↗ 90.0
LiveBench Coding: TypeScriptcoding 60%LiveBenchLiveBench subtask result; dataset edition 2026_06_25; publication time is HTTP Last-Modified of the score CSV, not the dataset edition or individual evaluation date [variant] 2026_06_25Published Oct 7, 2026 Retrieved Oct 9, 2026 · factual citation; Apache-2.0 code
Open source ↗ 60.0
Terminal-Bench 2.1coding 90.6%Official model cards via models.devLab-reported; metric pass@1; transcribed by MIT models.dev catalog; not independently evaluated [variant] reasoning_effort=100; DeepSeek Harness minimal mode; 2.1 Retrieved Oct 9, 2026 · factual citation; MIT transcription
Open source ↗ 90.6 8 settings
  • Setting 1 88%Official model cards via models.devLab-reported; metric pass@1; transcribed by MIT models.dev catalog; not independently evaluated [variant] reasoning_effort=100; Claude Code; 2.1 Retrieved Oct 9, 2026 · factual citation; MIT transcription
    Open source ↗
  • Setting 2 84.1%Official model cards via models.devLab-reported; metric pass@1; transcribed by MIT models.dev catalog; not independently evaluated [variant] reasoning_effort=100; Codex; 2.1 Retrieved Oct 9, 2026 · factual citation; MIT transcription
    Open source ↗
  • Setting 3 90.6%Official model cards via models.devLab-reported; metric pass@1; transcribed by MIT models.dev catalog; not independently evaluated [variant] reasoning_effort=100; DeepSeek Harness minimal mode; 2.1 Retrieved Oct 9, 2026 · factual citation; MIT transcription
    Open source ↗
  • Setting 4 85.8%Official model cards via models.devLab-reported; metric pass@1; transcribed by MIT models.dev catalog; not independently evaluated [variant] reasoning_effort=100; DeepSeek Harness PTC mode; 2.1 Retrieved Oct 9, 2026 · factual citation; MIT transcription
    Open source ↗
  • Setting 5 85.8%Official model cards via models.devLab-reported; metric pass@1; transcribed by MIT models.dev catalog; not independently evaluated [variant] reasoning_effort=100; DeepSeek Harness standard mode; 2.1 Retrieved Oct 9, 2026 · factual citation; MIT transcription
    Open source ↗
  • Setting 6 90.3%Official model cards via models.devLab-reported; metric pass@1; transcribed by MIT models.dev catalog; not independently evaluated [variant] reasoning_effort=100; mini-swe-agent; 2.1 Retrieved Oct 9, 2026 · factual citation; MIT transcription
    Open source ↗
  • Setting 7 85%Official model cards via models.devLab-reported; metric pass@1; transcribed by MIT models.dev catalog; not independently evaluated [variant] reasoning_effort=100; OpenCode; 2.1 Retrieved Oct 9, 2026 · factual citation; MIT transcription
    Open source ↗
  • Setting 8 86.1%Official model cards via models.devLab-reported; metric pass@1; transcribed by MIT models.dev catalog; not independently evaluated [variant] reasoning_effort=100; Pi; 2.1 Retrieved Oct 9, 2026 · factual citation; MIT transcription
    Open source ↗
About Terminal-Bench 2.1
Terminal-Bench 3.0coding 30%Official model cards via models.devLab-reported; metric pass@1; transcribed by MIT models.dev catalog; not independently evaluated [variant] reasoning_effort=100; DeepSeek Harness minimal mode; 3.0 Retrieved Oct 9, 2026 · factual citation; MIT transcription
Open source ↗ 30.0
Terminal-Bench 4.0coding 31.2%Official model cards via models.devLab-reported; metric pass@1; transcribed by MIT models.dev catalog; not independently evaluated [variant] reasoning_effort=100; DeepSeek Harness minimal mode; 4.0 Retrieved Oct 9, 2026 · factual citation; MIT transcription
Open source ↗ 31.2
LMArena Textpreference 1462.4 eloLMArena / ArenaPublished source fact [variant] text / overallPublished Oct 8, 2026 Retrieved Oct 9, 2026 · CC-BY-4.0
Open source ↗ 78.8

Normalization uses a fixed 0–100 scale for each unit, independent of other models. Compare evaluation conditions before reading a small gap as decisive. “Lab-reported” marks the provider's own published figure.

Details and sources

Open weights
YesHugging Face HubPublic Hub repo with weight files; gating/repo upload date is not release date Retrieved Oct 9, 2026 · factual metadata; model-specific licenses
Open source ↗
License
mitHugging Face HubPublished source fact Retrieved Oct 9, 2026 · factual metadata; model-specific licenses
Open source ↗
Input modalities
text, imagemodels.devPublished source fact Retrieved Oct 9, 2026 · MIT
Open source ↗
Popularity
#1 by usageOpenRouter rankingsPublished source fact; tokens processed; page default period; excludes catalog and endpoint statistics Retrieved Oct 9, 2026 · CC-BY-4.0
Open source ↗
First seen by SuperIndex
Oct 8, 2026
Coverage
55% of expected source weight

Reported (4)

  • ARC PrizeOct 8, 2026
  • Official model cards via models.devOct 8, 2026
  • LiveBenchOct 8, 2026
  • LMArena / ArenaOct 8, 2026

Awaiting (3)

  • Epoch AI Benchmarking25% of weight
  • Humanity’s Last Exam13% of weight
  • Terminal-Bench6% of weight

Confidence rises as pending sources publish. Some sources never cover some models, so confidence reaches 100% at 80% of expected weight.

What changed

All releases · RSS feed

Alerts on this device

What to be alerted about
RSS feed