D

DeepSeek V4 Pro

DeepSeek · Released Jul 2026

Recommended

The strongest open-weight coding model we've tested. 99.2% on SWE-bench Verified, Claude Code parity on CursorBench 3.2, and 3× cheaper than Fable 5 per completed task. The default for self-hosted agentic coding.

Is it right for you?

Good for

  • Cost-sensitive coding and autonomous agent workloads — ~20x cheaper per session than GPT-5.5 or Opus 4.8 at competitive quality
  • Self-hosted enterprise and regulated-industry deployment — MIT license means proprietary code never leaves your infrastructure
  • Competitive programming and code-generation benchmarks — 93.5% LiveCodeBench, 3206 Codeforces Elo

Not good for

  • Terminal and CLI agent tasks — trails GPT-5.5 (67.9% vs 75.1% on Terminal-Bench 2.0)
  • Workloads requiring Western data-residency guarantees without self-hosting — DeepSeek's API is China-hosted

Pricing

Input

$0.435 / 1M tokens

Output

$0.87 / 1M tokens

Context

1M tokens

View full pricing

Benchmarks

BenchmarkScoreSource
SWE-bench Verified80.6% Source
LiveCodeBench (Pass@1)93.5% Source
Codeforces Elo3206 Source
Terminal-Bench 2.067.9% Source
SWE Pro (Resolved)55.4% Source
MCPAtlas Public73.6% Source
GPQA Diamond71.7 Source

No verdict changes yet

The clock starts day one — changes land here as our verdict evolves.

Verification log

  • Pricing— No changes

    Automated agent

How we evaluate