DeepSeek

DeepSeek V4 Pro

Version 1.0 benchmark runs across the shared Superbash visual prompts.

API pricing

USD per million tokens · DeepSeek API peak rate
Input
$1.32
Output
$3.96
Cached input
$0.044
Context window
1M
Pricing details and specifications

Inspect the original visual runs

Open the generated scene to test it, or compare that prompt with another model.

10 runs

City Scroll Journey

Run date: 2026-08-17

Specialist test · evaluated separately

Generation provenance: unrecorded

Cinematic scroll journey testing AI-generated scene continuity, scroll-scrubbed camera motion, and art-directed landing craft.

Ember Glider

Run date: 2026-08-17

Generation provenance: unrecorded

Superbench 2.1 returning brief · original 1.x run protocol

Sunset gliding journey testing flight energy management, checkpoint flow, and atmospheric scene craft.

Helm's Deep

Run date: 2026-08-17

Generation provenance: unrecorded

Fortress siege scene testing scale, lighting, architecture, and cinematic atmosphere.

Hogwarts Broom Flight Simulator

Run date: 2026-08-17

Generation provenance: unrecorded

Broom-flight scene testing depth, motion cues, castle scale, and fantasy mood.

Low-Poly Tower Defense

Run date: 2026-08-17

Generation provenance: unrecorded

Superbench 2.1 returning brief · original 1.x run protocol

Diorama tower defense testing economy balance, wave design, placement rules, and combat readability.

Mechanical Watch Simulator

Run date: 2026-08-17

Generation provenance: unrecorded

Interactive watch movement testing mechanical legibility, accurate relative motion, and real-time 3D controls.

Neon Drift

Run date: 2026-08-17

Generation provenance: unrecorded

Synthwave time-trial racing testing drift physics, lap timing, ghost replay, and unlock progression.

Starfall Arena

Run date: 2026-08-17

Generation provenance: unrecorded

Neon arena survival testing wave escalation, upgrade builds, particle feedback, and boss design.

Stormwind Trebuchet Simulator

Run date: 2026-08-17

Generation provenance: unrecorded

Counterweight siege simulation testing coupled mechanics, trajectory prediction, projectile cameras, and interactive tuning.

Yingzao Fashi Assembly

Run date: 2026-08-17

Specialist test · evaluated separately

Generation provenance: unrecorded

Timber assembly scene testing structure, joinery, construction order, and material clarity.

Retired tests · 6 archived runs

These tests are no longer in the active suite. Their original results remain available for inspection.

How DeepSeek V4 Pro compares

Three relevant peers, side by side.

  • DeepSeek V4 Pro This model
  • DeepSeek V4 Flash Max
  • Claude Opus 4.6 Max
  • Gemini 3.1 Pro High

6 benchmarks · 4 models

MMLU-Pro

Knowledge · Higher is better

Source
DeepSeek V4 Pro87.5%
DeepSeek V4 Flash Max86.2%
Claude Opus 4.6 Max89.1%
Gemini 3.1 Pro High91.0%

GPQA Diamond

STEM reasoning · Higher is better

Source
DeepSeek V4 Pro90.1%
DeepSeek V4 Flash Max88.1%
Claude Opus 4.6 Max91.3%
Gemini 3.1 Pro High94.3%

LiveCodeBench

Code reasoning · Higher is better

Source
DeepSeek V4 Pro93.5%
DeepSeek V4 Flash Max91.6%
Claude Opus 4.6 Max88.8%
Gemini 3.1 Pro High91.7%

HMMT February 2026

Math · Higher is better

Source
DeepSeek V4 Pro95.2%
DeepSeek V4 Flash Max94.8%
Claude Opus 4.6 Max96.2%
Gemini 3.1 Pro HighNot reported

IMOAnswerBench

Math reasoning · Higher is better

Source
DeepSeek V4 Pro89.8%
DeepSeek V4 Flash Max88.4%
Claude Opus 4.6 MaxNot reported
Gemini 3.1 Pro HighNot reported

MRCR 1M

Long context · Higher is better

Source
DeepSeek V4 Pro83.5%
DeepSeek V4 Flash Max78.7%
Claude Opus 4.6 Max92.9%
Gemini 3.1 Pro High76.3%

Official benchmark profile

Published scores outside our visual tests

DeepSeek sourceApril 2026Source report →
Knowledge

MMLU-Pro

87.5%
STEM reasoning

GPQA Diamond

90.1%
Code reasoning

LiveCodeBench

93.5%
Full official benchmark table12 scores and peer comparisons
BenchmarkAreaScoreComparison
MMLU-ProKnowledge87.5%
DeepSeek V4 Pro87.5%
DeepSeek V4 Flash Max86.2%
Claude Opus 4.6 Max89.1%
Gemini 3.1 Pro High91.0%
GPQA DiamondSTEM reasoning90.1%
DeepSeek V4 Pro90.1%
Claude Opus 4.6 Max91.3%
DeepSeek V4 Flash Max88.1%
GPT-5.4 xHigh93.0%
LiveCodeBenchCode reasoning93.5%
DeepSeek V4 Pro93.5%
Gemini 3.1 Pro High91.7%
DeepSeek V4 Flash Max91.6%
Kimi K2.6 Thinking89.6%
HMMT February 2026Math95.2%
DeepSeek V4 Pro95.2%
DeepSeek V4 Flash Max94.8%
Claude Opus 4.6 Max96.2%
GPT-5.4 xHigh97.7%
IMOAnswerBenchMath reasoning89.8%
DeepSeek V4 Pro89.8%
DeepSeek V4 Flash Max88.4%
GPT-5.4 xHigh91.4%
Kimi K2.6 Thinking86.0%
MRCR 1MLong context83.5%
DeepSeek V4 Pro83.5%
DeepSeek V4 Flash Max78.7%
Gemini 3.1 Pro High76.3%
Claude Opus 4.6 Max92.9%
Terminal-Bench 2.0Agentic coding67.9%
DeepSeek V4 Pro67.9%
Gemini 3.1 Pro High68.5%
Kimi K2.6 Thinking66.7%
GPT-5.4 xHigh75.1%
SWE-bench VerifiedAgentic coding80.6%
DeepSeek V4 Pro80.6%
Gemini 3.1 Pro High80.6%
Claude Opus 4.6 Max80.8%
DeepSeek V4 Flash Max79.0%
SWE-bench ProAgentic coding55.4%
DeepSeek V4 Pro55.4%
GPT-5.4 xHigh57.7%
DeepSeek V4 Flash Max52.6%
GLM 5.1 Thinking58.4%
BrowseCompWeb research83.4%
DeepSeek V4 Pro83.4%
Claude Opus 4.6 Max83.7%
Gemini 3.1 Pro High85.9%
DeepSeek V4 Flash Max73.2%
MCP AtlasTool use73.6%
DeepSeek V4 Pro73.6%
Claude Opus 4.6 Max73.8%
GLM 5.1 Thinking71.8%
Gemini 3.1 Pro High69.2%
ToolathlonTool use51.8%
DeepSeek V4 Pro51.8%
Kimi K2.6 Thinking50.0%
GPT-5.4 xHigh54.6%
DeepSeek V4 Flash Max47.8%
Sources, pricing details and methodology

Verified model facts

Model specifications

Pricing source · checked 2026-09-08 ↗
Status
Superseded in this ranking
API model ID
deepseek-v4-pro
Context
1M
Max output
384K
License
Not publicly verified
Access route
Not publicly verified
Modalities
Not publicly verified
Hardware
Not publicly verified
DeepSeek API peak rate · USD / 1M tokens
$1.32 input · $0.044 cached input · $3.96 output

Peak hours: Monday–Friday, 01:00–04:00 and 06:00–10:00 UTC. All other hours use half-price off-peak rates: $0.66 input, $0.022 cached input, and $1.98 output per 1M tokens.

Benchmark sources

DeepSeek V4 Pro activates 49B of 1.6T parameters with a 1M-token context. These are the official V4-Pro Max results, which use the largest published reasoning budget and differ from the non-thinking and high settings.

DeepSeek-V4: Towards Highly Efficient Million-Token Context Intelligence

Evaluation settings

  • Exact match at V4-Pro Max reasoning effort.
  • Pass@1 at V4-Pro Max reasoning effort.
  • Mean matching rate at the 1M-token setting.
  • Accuracy in the official DeepSeek comparison table.
  • Resolved rate in the official DeepSeek comparison table.
  • Pass@1 in the official DeepSeek comparison table.

Edition coverage

Recorded means a published artifact, not a behavioral pass. Historical returning results retain their original protocol.

Superbench 2.1: 2/6 recorded · 0 pending · 4 not run
  • 15 Ember Glider · repository-run · Recorded
  • 14 Low-Poly Tower Defense · repository-run · Recorded
  • 21 Skyhook Cargo · single-prompt · Not run — Historical returning-brief artifacts only; the four new 2.1 contracts have not been published for this model version.
  • 22 Concourse Divide · single-prompt · Not run — Historical returning-brief artifacts only; the four new 2.1 contracts have not been published for this model version.
  • 23 Lightkeeper's Atelier · single-prompt · Not run — Historical returning-brief artifacts only; the four new 2.1 contracts have not been published for this model version.
  • 24 Wrecklight Salvage · single-prompt · Not run — Historical returning-brief artifacts only; the four new 2.1 contracts have not been published for this model version.