DeepSeek

DeepSeek V4.1 Flash

Version 4.1 benchmark runs across the shared Superbash visual prompts.

API pricing

USD per million tokens · DeepSeek API peak rate
Input
$0.3
Output
$1.2
Cached input
$0.006
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: 2026-09-22

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: 2026-09-22

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: 2026-09-22

Generation provenance: unrecorded

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

Hogwarts Broom Flight Simulator

Run date: 2026-09-22

Generation provenance: unrecorded

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

Low-Poly Tower Defense

Run date: 2026-09-22

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: 2026-09-22

Generation provenance: unrecorded

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

Neon Drift

Run date: 2026-09-22

Generation provenance: unrecorded

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

Starfall Arena

Run date: 2026-09-22

Generation provenance: unrecorded

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

Stormwind Trebuchet Simulator

Run date: 2026-09-22

Generation provenance: unrecorded

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

Yingzao Fashi Assembly

Run date: 2026-09-22

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 V4.1 Flash compares

2 peers with published comparison scores.

  • DeepSeek V4.1 Flash This model
  • DeepSeek V4 Flash
  • DeepSeek V4 Pro

6 benchmarks · 3 models

GPQA Diamond

STEM reasoning · Higher is better

Source
DeepSeek V4.1 Flash90.9%
DeepSeek V4 Flash89.9%
DeepSeek V4 Pro92.4%

Codeforces

Competitive programming · Higher is better

Source
DeepSeek V4.1 Flash3471
DeepSeek V4 Flash3289
DeepSeek V4 Pro3348

Terminal-Bench 2.1

Agentic coding · Higher is better

Source
DeepSeek V4.1 Flash90.6%
DeepSeek V4 Flash82.7%
DeepSeek V4 Pro87.9%

DeepSWE v1.1

Agentic coding · Higher is better

Source
DeepSeek V4.1 Flash74.2%
DeepSeek V4 Flash54.4%
DeepSeek V4 Pro62.7%

CyberGym

Cybersecurity agent · Higher is better

Source
DeepSeek V4.1 Flash88.1%
DeepSeek V4 Flash76.7%
DeepSeek V4 Pro83.3%

SEC-Bench Pro

Security coding · Higher is better

Source
DeepSeek V4.1 Flash62.8%
DeepSeek V4 Flash30.9%
DeepSeek V4 Pro56.4%

Official benchmark profile

Published scores outside our visual tests

DeepSeek sourceSeptember 2026Source report →
STEM reasoning

GPQA Diamond

90.9%
Competitive programming

Codeforces

3471
Agentic coding

Terminal-Bench 2.1

90.6%
Full official benchmark table10 scores and peer comparisons
BenchmarkAreaScoreComparison
GPQA DiamondSTEM reasoning90.9%
DeepSeek V4.1 Flash90.9%
DeepSeek V4 Flash89.9%
DeepSeek V4 Pro92.4%
CodeforcesCompetitive programming3471
Terminal-Bench 2.1Agentic coding90.6%
DeepSeek V4.1 Flash90.6%
DeepSeek V4 Pro87.9%
DeepSeek V4 Flash82.7%
DeepSWE v1.1Agentic coding74.2%
DeepSeek V4.1 Flash74.2%
DeepSeek V4 Pro62.7%
DeepSeek V4 Flash54.4%
CyberGymCybersecurity agent88.1%
DeepSeek V4.1 Flash88.1%
DeepSeek V4 Pro83.3%
DeepSeek V4 Flash76.7%
SEC-Bench ProSecurity coding62.8%
DeepSeek V4.1 Flash62.8%
DeepSeek V4 Pro56.4%
DeepSeek V4 Flash30.9%
HLE with toolsTool-using reasoning63.9%
DeepSeek V4.1 Flash63.9%
DeepSeek V4 Pro60.0%
DeepSeek V4 Flash51.5%
AutomationBenchAgentic automation54.8%
DeepSeek V4.1 Flash54.8%
DeepSeek V4 Pro43.2%
DeepSeek V4 Flash37.7%
Agent's Last ExamAgentic reasoning31.8%
DeepSeek V4.1 Flash31.8%
DeepSeek V4 Pro25.7%
DeepSeek V4 Flash25.2%
ZeroBench-main with toolsVisual agent49.0%
Sources, pricing details and methodology

Verified model facts

Model specifications

Pricing source · checked 2026-09-21 ↗
Status
Superseded in this ranking
API model ID
deepseek-flash
Context
1M
Max output
384K
License
MIT
Access route
September 2026 DeepSeek API identity. At launch, the deepseek-flash model ID and legacy V4 aliases routed here; the alias advanced to Flash-1017 on October 17. This record preserves the September visual runs.
Modalities
text input, image input, text output
Hardware
Not publicly verified
DeepSeek API peak rate · USD / 1M tokens
$0.3 input · $0.006 cached input · $1.2 output

Peak hours are Monday-Friday, 01:00-04:00 and 06:00-10:00 UTC. Off-peak rates are half: $0.15 input, $0.003 cached input, and $0.60 output per 1M tokens.

Benchmark sources

DeepSeek V4.1 Flash is a 552B-parameter multimodal MoE model with a 1M-token context window. The provider reports 8B active parameters during prefill and 16B during decoding. Scores below are the provider's maximum-reasoning results; agent benchmarks use different harnesses and are not a universal leaderboard.

DeepSeek-V4.1-Flash: Pushing the Limits of KV Cache Compression

These are DeepSeek-reported results, not Superbash visual benchmark runs. The provider reports maximum reasoning effort (reasoning_effort=100), its named harnesses, and a 1M-token context limit for the code-agent evaluations.

Evaluation settings

  • Pass@1 at maximum reasoning effort; DeepSeek-reported.
  • Rating at maximum reasoning effort; DeepSeek-reported.
  • Pass@1 using DeepSeek Harness Minimal and a 1M-token context; DeepSeek-reported.
  • Resolved rate using mini-SWE and a 1M-token context; DeepSeek-reported.
  • Pass@1 using the Claude Code harness; DeepSeek-reported.
  • Pass@1 using the official scaffold; DeepSeek-reported.
  • Pass@5 using the Claude Code harness and a 512K-token context; DeepSeek-reported.

Archived September visual-run methodology and validation records

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.