Frontier BenchSI model benchmarks · Jul 2026
ModelsCompareLabsMethodology
All models
D

DeepSeek V4 Pro

DeepSeek

Best open-weight coder. MIT license, 1M context, 49B active MoE.

Open-weightMIT1M contextReleased Apr 2026textcode

API price · per 1M tokens

$0.435/$0.87

input / output

≈ $0.54/M blended

Benchmarks

Bar shows this model's score; tick shows the best score across all tracked models for context.

SWE-bench VerifiedReal GitHub issues, human-verified patches. The standard coding-agent benchmark.
83.7/ best 90.1
SWE-bench Pro2,294 harder real-world GitHub issues. Where the frontier separates now.
55.4/ best 69.2
Terminal-Bench 2.xAgentic command-line workflows — planning, iteration, tool coordination.
67.9/ best 84.0
GPQA DiamondGoogle-proof graduate-level science Q&A. General reasoning proxy.
85.0/ best 94.4
MMLU-ProMulti-task language understanding across 120+ academic subjects.
82.0/ best 93.0
HumanEval+Functional code generation from docstrings.
91.0/ best 96.0
MATH-500Competition-style math problems.
94.2/ best 97.8
Humanity's Last ExamHardest reasoning benchmark (with tools). ~50% = near frontier.
38.0/ best 57.9
AIME 2025American Invitational Mathematics Examination.
88.0/ best 96.5

Other DeepSeek models

D

DeepSeek V4 Flash

Open-weight

≈$0.18/M

Built by GLM-5.2 · Scores are directional, drawn from public mid-2026 benchmarks.

SWE-bench, GPQA, MMLU, AIME, HLE etc. are trademarks of their respective owners.