All models
DeepSeek

DeepSeek V4

frontierOpen Source

DeepSeek

DeepSeek's latest flagship MoE model with 1.6T parameters and 49B active. Dramatically expands context to 1M tokens with optional thinking mode — still at a fraction of the cost of Western labs.

Official page
open-sourcelong-contextcost-effectivemoe
Context
1.048576M
Input / 1M
$0.43
Output / 1M
$0.87
Knowledge cutoff
2025-06

Capabilities

Code5/5
Reasoning5/5
Multilingual3/5
Safety2/5
Speed3/5

Modalities & Features

Vision (image input) No
Audio No
Video No
Function / tool calling Yes
JSON mode Yes
Extended thinking Yes

Context & Output

Context window1.048576M
Max output tokens32.768K
Knowledge cutoff2025-06
Release date2026-03

API & Access

Open source Yes
Streaming Yes
Batch API No
Prompt caching Yes
Fine-tuning No

Benchmarks

MMLUGeneral knowledge
87.5%
GPQA DiamondHard science Q&A
90.1%
MATHMath problem solving
88%
AIME 2025Advanced math
N/A
GSM8KGrade-school math
N/A
HumanEvalCode generation
N/A
LiveCodeBenchLive coding problems
93.5%
SWE-bench VerifiedReal-world software
80.6%
HellaSwagCommonsense inference
N/A

Source: provider technical reports & independent evaluations. N/A = not yet published for this model.

Pricing (per 1M tokens)

Input$0.43
Output$0.87
Consumer plan
DeepSeek Chat
Free
Prices shown are list prices and may vary by provider plan. Always check official documentation for the latest pricing.