LLMcompare

Search

Search for a command to run...

A · open · DeepSeek

DeepSeek V3

DeepSeek

B · open · DeepSeek

DeepSeek V4 Flash

DeepSeek

1:12

overall capability wins (0 ties across 13 scored). DeepSeek V4 Flash leads overall.

  • Reasoning0:5
  • Coding0:3
  • Arena0:1
  • Tool use0:1
  • Composite indices0:2
  • Pricing0:3
  • Speed0:2
  • Specs1:2

Verdict

DeepSeek V4 Flash leads coding (widest gap: +44.9 on TermBench); DeepSeek V4 Flash ranks higher on LMArena (+77 Elo); DeepSeek V4 Flash offers ~7.8x larger context (1M vs 128K).

  • DeepSeek V4 Flash leads coding (widest gap: +44.9 on TermBench)
  • DeepSeek V4 Flash ranks higher on LMArena (+77 Elo)
  • DeepSeek V4 Flash offers ~7.8x larger context (1M vs 128K)
  • DeepSeek V4 Flash allows ~46.9x more max output (384K vs 8K)
  • DeepSeek V4 Flash is ~2.7x cheaper on a blended token basis than DeepSeek V3

Cost for 1M in + 250K out

Illustrative chat workload at primary-provider list prices.

DeepSeek V3
$0.55
DeepSeek V4 Flash
$0.21

DeepSeek V4 Flash is about 2.6x cheaper on this workload.

Price per 1M tokens

Full comparison

Organization

DeepSeek

DeepSeek

Family

DeepSeek

DeepSeek

License

MIT

MIT

Open weights

Yes

Yes

Release

Dec 26, 2024

Apr 24, 2026

Knowledge cutoff

-

-

API / provider

DeepSeek

DeepSeek

Modalities

text → text

text → text

Specs

Context window

128K

1M

B +872K

Max output

8K

384K

B +376K

Parameters

671B (37B act.)

284B (13B act.)

A +387

Pricing

Input $/1M

$0.27

$0.14

B +$0.13

Output $/1M

$1.10

$0.28

B +$0.82

Blended $/1M (3∶1)

$0.48

$0.18

B +$0.30

Speed

tok/s

80

120

B +40

TTFT (s)

0.4

0.25

B +0.15

Reasoning

MMLU-Pro

75.9

—

—

GPQA Diamond

55.7

89.4

B +33.7 pts

Humanity's Last Exam

2.9

34.8

B +31.9 pts

AIME 2025

26.0

—

—

MATH-500

90.2

—

—

Humanity's Last Exam (with tools)

—

—

—

AA-Omniscience Accuracy

25.5

36.8

B +11.3 pts

AA-LCR v1.1

29.3

74.3

B +45.0 pts

CritPt

—

7.1

—

MMMU-Pro

—

—

—

IFBench

34.8

79.2

B +44.4 pts

Chartography

—

—

—

Chartography (With Tools)

—

—

—

Coding

SWE-bench Verified

—

—

—

SWE-bench Pro

—

—

—

SWE-bench Multilingual

—

—

—

LiveCodeBench

49.6

—

—

Terminal-Bench 2.1

16.9

61.8

B +44.9 pts

Aider Polyglot

55.1

—

—

Terminal-Bench 3

—

—

—

BigCodeBench

50.0

—

—

SciCode

35.8

45.3

B +9.5 pts

CursorBench

—

—

—

SWE-Rebench

23.3

38.4

B +15.1 pts

NL2Repo-Bench

—

39.4

—

DeepSWE

—

7.3

—

WebDev Arena

—

1,580

—

Terminal-Bench 4.0

—

—

—

CursorBench 4.0

—

—

—

FrontierCode v1.1 (Main)

—

—

—

Arena

LMArena Elo

1,358

1,436

B +77 Elo

DeepSeek V3

Context
128K / 8K out
Parameters
671B (37B act.)
Price
$0.27 / $1.10
Speed
80 tok/s · 0.4s TTFT
Modalities
text → text
License
MIT

Original DeepSeek V3 MoE - the release that reset open-weight price/performance expectations.

DeepSeek V4 Flash

Context
1M / 384K out
Parameters
284B (13B act.)
Price
$0.14 / $0.28
Speed
120 tok/s · 0.25s TTFT
Modalities
text → text
License
MIT

Budget DeepSeek V4 API tier (Apr 2026 preview). Superseded by DeepSeek-V4-Flash-0731 (public beta), which re-post-trains the same architecture and reports far stronger agent benchmarks.

Benchmark charts

Winner bars are emphasized. Per-benchmark deltas sit above each chart.

Reasoning & knowledge

GPQA: B +33.7HLE: B +31.9AA-LCR: B +45.0Omniscience Acc.: B +11.3IFBench: B +44.4

Coding

TermBench: B +44.9SciCode: B +9.5SWE-Rebench: B +15.1

Tool use & function calling

τ³-Banking: B +26.2

Arena

Arena Elo: B +77

Composite indices

AA Index: B +16.1AA Index v4.3: B +16.1

Pick a different pair · Back to catalog