vs
5:1
overall capability wins (2 ties across 8 scored). DeepSeek V4.1 Flash leads overall.
- Reasoning3:1
- Coding1:0
- Composite indices1:1
- Pricing3:0
- Specs1:0
Verdict
DeepSeek V4.1 Flash leads coding (widest gap: +1.6 on SciCode); DeepSeek V4.1 Flash is ~1.3x cheaper on a blended token basis than DeepSeek V4 Flash 0731.
- DeepSeek V4.1 Flash leads coding (widest gap: +1.6 on SciCode)
- DeepSeek V4.1 Flash is ~1.3x cheaper on a blended token basis than DeepSeek V4 Flash 0731
Cost for 1M in + 250K out
Illustrative chat workload at primary-provider list prices.
- DeepSeek V4.1 Flash
- $0.60
- DeepSeek V4 Flash 0731
- $0.77
DeepSeek V4.1 Flash is about 1.3x cheaper on this workload.
Price per 1M tokens
Full comparison
Organization
DeepSeek
DeepSeek
Family
DeepSeek
DeepSeek
License
MIT
MIT
Open weights
Yes
Yes
Release
Sep 10, 2026
Jul 31, 2026
Knowledge cutoff
-
-
API / provider
DeepSeek
DeepSeek
Modalities
text, image → text
text → text
Specs
Context window
1M
1M
Max output
384K
384K
Parameters
552B (16B act.)
304B
Pricing
Input $/1M
$0.30
$0.44
Output $/1M
$1.20
$1.32
Blended $/1M (3∶1)
$0.52
$0.66
Speed
tok/s
-
-
TTFT (s)
-
-
Reasoning
MMLU-Pro
—
—
GPQA Diamond
—
90.8
Humanity's Last Exam
39.2
38.6
AIME 2025
—
—
MATH-500
—
—
Humanity's Last Exam (with tools)
—
—
AA-Omniscience Accuracy
46.4
40.4
AA-LCR v1.1
84.0
79.7
CritPt
14.3
16.6
MMMU-Pro
77.0
—
IFBench
—
—
Chartography
—
—
Chartography (With Tools)
—
—
Coding
SWE-bench Verified
—
—
SWE-bench Pro
—
—
SWE-bench Multilingual
—
—
LiveCodeBench
—
—
Terminal-Bench 2.1
—
78.7
Aider Polyglot
—
—
Terminal-Bench 3
—
—
BigCodeBench
—
—
SciCode
51.9
50.3
CursorBench
—
—
SWE-Rebench
—
—
NL2Repo-Bench
—
54.2
DeepSWE
—
54.4
WebDev Arena
1,620
—
Terminal-Bench 4.0
—
—
CursorBench 4.0
—
—
FrontierCode v1.1 (Main)
—
—
Arena
LMArena Elo
—
—
| Metric | DeepSeek V4.1 Flash | DeepSeek V4 Flash 0731 | Delta |
|---|---|---|---|
| Identity | |||
| Organization | DeepSeek | DeepSeek | — |
| Family | DeepSeek | DeepSeek | — |
| License | MIT | MIT | — |
| Open weights | Yes | Yes | — |
| Release | Sep 10, 2026 | Jul 31, 2026 | — |
| Knowledge cutoff | - | - | — |
| API / provider | DeepSeek | DeepSeek | — |
| Modalities | text, image → text | text → text | — |
| Specs | |||
| Context window | 1M | 1M | tie |
| Max output | 384K | 384K | tie |
| Parameters | 552B (16B act.) | 304B | A +248 |
| Pricing | |||
| Input $/1M | $0.30 | $0.44 | A +$0.14 |
| Output $/1M | $1.20 | $1.32 | A +$0.12 |
| Blended $/1M (3∶1) | $0.52 | $0.66 | A +$0.14 |
| Speed | |||
| tok/s | - | - | — |
| TTFT (s) | - | - | — |
| Reasoning | |||
| MMLU-Pro | — | — | — |
| GPQA Diamond | — | 90.8 | — |
| Humanity's Last Exam | 39.2 | 38.6 | A +0.6 pts |
| AIME 2025 | — | — | — |
| MATH-500 | — | — | — |
| Humanity's Last Exam (with tools) | — | — | — |
| AA-Omniscience Accuracy | 46.4 | 40.4 | A +6.0 pts |
| AA-LCR v1.1 | 84.0 | 79.7 | A +4.3 pts |
| CritPt | 14.3 | 16.6 | B +2.3 pts |
| MMMU-Pro | 77.0 | — | — |
| IFBench | — | — | — |
| Chartography | — | — | — |
| Chartography (With Tools) | — | — | — |
| Coding | |||
| SWE-bench Verified | — | — | — |
| SWE-bench Pro | — | — | — |
| SWE-bench Multilingual | — | — | — |
| LiveCodeBench | — | — | — |
| Terminal-Bench 2.1 | — | 78.7 | — |
| Aider Polyglot | — | — | — |
| Terminal-Bench 3 | — | — | — |
| BigCodeBench | — | — | — |
| SciCode | 51.9 | 50.3 | A +1.6 pts |
| CursorBench | — | — | — |
| SWE-Rebench | — | — | — |
| NL2Repo-Bench | — | 54.2 | — |
| DeepSWE | — | 54.4 | — |
| WebDev Arena | 1,620 | — | — |
| Terminal-Bench 4.0 | — | — | — |
| CursorBench 4.0 | — | — | — |
| FrontierCode v1.1 (Main) | — | — | — |
| Arena | |||
| LMArena Elo | — | — | — |
DeepSeek V4.1 Flash
- Context
- 1M / 384K out
- Parameters
- 552B (16B act.)
- Price
- $0.30 / $1.20 peak · $0.15 / $0.60 off-peak
- Speed
- —
- Modalities
- text, image → text
- License
- MIT
New default DeepSeek Flash tier (Sep 10, 2026), the smallest model in a new Causal Encoder-Decoder architecture family: 552B total parameters activating 8B (prefill) / 16B (decode) per token, MIT-licensed weights on Hugging Face. Replaces DeepSeek-V4-Flash and V4-Flash-Vision-Exp, which are retired but still routed to this model for compatibility. DeepSeek's own reported numbers (GPQA Diamond 90.9, Terminal-Bench 2.1 90.6, text-only HLE 39.1) are self-measured and not directly comparable to this catalog's Artificial-Analysis-sourced cells. API pricing: $0.30 in / $1.20 out peak, $0.15 in / $0.60 out off-peak (per 1M tokens); off-peak cache-hit input is $0.003.
DeepSeek V4 Flash 0731
- Context
- 1M / 384K out
- Parameters
- 304B
- Price
- $0.44 / $1.32 peak · $0.22 / $0.66 off-peak
- Speed
- —
- Modalities
- text → text
- License
- MIT
Official DeepSeek-V4-Flash release (Jul 31, 2026 public beta) superseding the preview - same base architecture plus a DSpark speculative-decoding module (304B total), MIT weights on Hugging Face, and agent benchmarks well above V4-Pro-Preview (Terminal-Bench 2.1 82.7, NL2Repo 54.2). Peak/off-peak API pricing took effect Aug 16, 2026: $0.44 in / $1.32 out peak, $0.22 in / $0.66 out off-peak (per 1M tokens).
Benchmark charts
Winner bars are emphasized. Per-benchmark deltas sit above each chart.