Model

DeepSeek-V4 Pro NO.28

Open MIT flagship — 1.6T-parameter MoE with native vision and a 1M-token context at throwaway prices.

DeepSeek · DeepSeek · Open · Open weights

Specification

Fig.01DeepSeek — DeepSeek

CHINA

Context window
1Mtokens
Max output
384Ktokens
Input price
$0.43/ 1M
Output price
$0.87/ 1M
Cached input
$0.04/ 1M
Throughput
Class
Open
Modalities
Text, Image
Weights
Open
Parameters
1600B
License
MIT
Knowledge cutoff
Jan 2026
Released
Apr 2026

Overview

DeepSeek-V4 Pro is DeepSeek's open-weight flagship (1.6T-parameter MoE, 49B active, MIT) and the headline of the V4 generation. It expands the context window 8x to 1M tokens, adds native image input, and introduces a refined DeepSeek Sparse Attention (a CSA/HCA hybrid) that cuts long-context compute and KV-cache footprint dramatically. DeepSeek advertises a chart-topping 80.6% on SWE-bench Verified, but independent contamination-resistant runs land far lower (~58%) and it trails Claude Opus on SWE-bench Pro, so it is a spectacular-value workhorse rather than a frontier-reliable autonomous coder. Its API speaks both OpenAI and Anthropic formats, dropping into Claude Code without a proxy.

Strengths

  • Among the lowest frontier-class prices in the industry
  • 1M-token context via efficient DeepSeek Sparse Attention
  • Native image input (new to the V4 generation)
  • Very cheap cached input; MIT open weights
  • Strong math and general reasoning value

Trade-offs

  • Vendor benchmarks overstate real-world coding reliability
  • Agentic-coding reliability below the Claude/GPT frontier
  • Vision quality trails dedicated multimodal flagships
  • Large 1.6T MoE — serving cost and latency

Fit

Best for

  • High-volume, cost-sensitive workloads at scale
  • Long-context document and repo processing
  • Math and reasoning tasks
  • Budget Claude Code-compatible backend
  • Self-hosted / on-prem deployments

Not ideal for

  • Unsupervised long-horizon autonomous coding
  • Mission-critical frontier reasoning

Capabilities

Ten axes, normalized 0–100 and scored the same way across the whole catalog — so a 78 here means what a 78 means anywhere else on the bench.

Fig.02Capability radar

0 — 100

Reasoning
85
Coding
73
Math
91
Writing
81
Knowledge
85
Speed
60
Agentic
69
Vision
66
Multilingual
83
Long Context
87

Benchmarks

Public results, with independent third-party runs marked. Bars normalize percentages against 100 and Elo ratings against a 1500 ceiling.

Independent index

Artificial Analysis Intelligence Index

Composite of ~9–10 independent evals · Artificial Analysis

54/ 100

BenchmarkResult
Artificial Analysis Intelligence IndexIndependentGeneral · Artificial Analysis
54%
SWE-bench VerifiedIndependentCoding · SWE-bench
80.6%
SWE-bench ProIndependentCoding · SWE-bench Pro
55.4%
GPQA DiamondIndependentReasoning · Independent aggregators
83%
AIME 2025Math · Vendor-reported
93%
LiveCodeBenchIndependentCoding · LiveCodeBench
78%
MMLU-ProKnowledge · Vendor-reported
86%
DeepSWEIndependentCoding · Datacurve
8%
LMArena EloIndependentGeneral · LMArena
1462Elo

Alternatives

Models in roughly the same class — the ones worth weighing against this record.

Index

All models/EDUCATION