vs
1:2
overall capability wins (0 ties across 3 scored). DeepSeek V4 Flash Vision Exp leads overall.
- Pricing3:0
- Specs1:2
Verdict
DeepSeek V4 Flash Vision Exp offers ~7.8x larger context (1M vs 128K); DeepSeek V4 Flash Vision Exp allows ~46.9x more max output (384K vs 8K); DeepSeek V3 ships open weights (MIT).
- DeepSeek V4 Flash Vision Exp offers ~7.8x larger context (1M vs 128K)
- DeepSeek V4 Flash Vision Exp allows ~46.9x more max output (384K vs 8K)
- DeepSeek V3 ships open weights (MIT)
- DeepSeek V3 is ~1.4x cheaper on a blended token basis than DeepSeek V4 Flash Vision Exp
Cost for 1M in + 250K out
Illustrative chat workload at primary-provider list prices.
- DeepSeek V3
- $0.55
- DeepSeek V4 Flash Vision Exp
- $0.77
DeepSeek V3 is about 1.4x cheaper on this workload.
Price per 1M tokens
Full comparison
Organization
DeepSeek
DeepSeek
Family
DeepSeek
DeepSeek
License
MIT
Proprietary
Open weights
Yes
No
Release
Dec 26, 2024
Aug 21, 2026
Knowledge cutoff
-
-
API / provider
DeepSeek
DeepSeek
Modalities
text → text
text, image → text
Specs
Context window
128K
1M
Max output
8K
384K
Parameters
671B (37B act.)
304B
Pricing
Input $/1M
$0.27
$0.44
Output $/1M
$1.10
$1.32
Blended $/1M (3∶1)
$0.48
$0.66
Speed
tok/s
80
-
TTFT (s)
0.4
-
Reasoning
MMLU-Pro
75.9
—
GPQA Diamond
55.7
—
Humanity's Last Exam
2.9
—
AIME 2025
26.0
—
MATH-500
90.2
—
Humanity's Last Exam (with tools)
—
—
AA-Omniscience Accuracy
25.5
—
AA-LCR v1.1
29.3
—
CritPt
—
—
MMMU-Pro
—
—
IFBench
34.8
—
Chartography
—
—
Chartography (With Tools)
—
—
Coding
SWE-bench Verified
—
—
SWE-bench Pro
—
—
SWE-bench Multilingual
—
—
LiveCodeBench
49.6
—
Terminal-Bench 2.1
16.9
—
Aider Polyglot
55.1
—
Terminal-Bench 3
—
—
BigCodeBench
50.0
—
SciCode
35.8
—
CursorBench
—
—
SWE-Rebench
23.3
—
NL2Repo-Bench
—
—
DeepSWE
—
—
WebDev Arena
—
—
Terminal-Bench 4.0
—
—
CursorBench 4.0
—
—
FrontierCode v1.1 (Main)
—
—
Arena
LMArena Elo
1,358
—
| Metric | DeepSeek V3 | DeepSeek V4 Flash Vision Exp | Delta |
|---|---|---|---|
| Identity | |||
| Organization | DeepSeek | DeepSeek | — |
| Family | DeepSeek | DeepSeek | — |
| License | MIT | Proprietary | — |
| Open weights | Yes | No | — |
| Release | Dec 26, 2024 | Aug 21, 2026 | — |
| Knowledge cutoff | - | - | — |
| API / provider | DeepSeek | DeepSeek | — |
| Modalities | text → text | text, image → text | — |
| Specs | |||
| Context window | 128K | 1M | B +872K |
| Max output | 8K | 384K | B +376K |
| Parameters | 671B (37B act.) | 304B | A +367 |
| Pricing | |||
| Input $/1M | $0.27 | $0.44 | A +$0.17 |
| Output $/1M | $1.10 | $1.32 | A +$0.22 |
| Blended $/1M (3∶1) | $0.48 | $0.66 | A +$0.18 |
| Speed | |||
| tok/s | 80 | - | — |
| TTFT (s) | 0.4 | - | — |
| Reasoning | |||
| MMLU-Pro | 75.9 | — | — |
| GPQA Diamond | 55.7 | — | — |
| Humanity's Last Exam | 2.9 | — | — |
| AIME 2025 | 26.0 | — | — |
| MATH-500 | 90.2 | — | — |
| Humanity's Last Exam (with tools) | — | — | — |
| AA-Omniscience Accuracy | 25.5 | — | — |
| AA-LCR v1.1 | 29.3 | — | — |
| CritPt | — | — | — |
| MMMU-Pro | — | — | — |
| IFBench | 34.8 | — | — |
| Chartography | — | — | — |
| Chartography (With Tools) | — | — | — |
| Coding | |||
| SWE-bench Verified | — | — | — |
| SWE-bench Pro | — | — | — |
| SWE-bench Multilingual | — | — | — |
| LiveCodeBench | 49.6 | — | — |
| Terminal-Bench 2.1 | 16.9 | — | — |
| Aider Polyglot | 55.1 | — | — |
| Terminal-Bench 3 | — | — | — |
| BigCodeBench | 50.0 | — | — |
| SciCode | 35.8 | — | — |
| CursorBench | — | — | — |
| SWE-Rebench | 23.3 | — | — |
| NL2Repo-Bench | — | — | — |
| DeepSWE | — | — | — |
| WebDev Arena | — | — | — |
| Terminal-Bench 4.0 | — | — | — |
| CursorBench 4.0 | — | — | — |
| FrontierCode v1.1 (Main) | — | — | — |
| Arena | |||
| LMArena Elo | 1,358 | — | — |
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 Vision Exp
- Context
- 1M / 384K out
- Parameters
- 304B
- Price
- $0.44 / $1.32 peak · $0.22 / $0.66 off-peak
- Speed
- —
- Modalities
- text, image → text
- License
- Proprietary
Experimental API-only vision variant of DeepSeek-V4-Flash-0731 (released Aug 21, 2026), adding image input (up to 600 images/request) while matching the text model on agents, reasoning, and world knowledge. Same rate card as V4-Flash: $0.44 in / $1.32 out peak, $0.22 / $0.66 off-peak, cache-hit $0.007 (per 1M). Benchmarks are DeepSeek's self-reported Harness numbers — close to but trailing Opus 4.8 on most rows (Terminal-Bench 2.1 83.9).
Benchmark charts
Winner bars are emphasized. Per-benchmark deltas sit above each chart.
These two models have no overlapping published benchmarks in our dataset. Compare specs and pricing instead.