DeepSeek V4-Flash vs GPT-5.4 mini vs Claude Haiku 4.5 vs Qwen3-235B-A22B

Best cell per row highlighted. Null means undisclosed — never counted as zero.

DeepSeek
DeepSeek V4-Flashretired
8.7
AI Panel
OpenAI
GPT-5.4 miniGA
9.0
AI Panel
Anthropic
Claude Haiku 4.5GA
8.4
AI Panel
Alibaba Cloud
Qwen3-235B-A22BGA
8.5
AI Panel

Identity & lifecycle

ProviderDeepSeekOpenAIAnthropicAlibaba Cloud
Family / tierDeepSeek V4 · FlashGPT-5 · miniClaude 4 · HaikuQwen3 · Large
StatusretiredGAGAGA
Released2026-04-242026-03-172025-10-152025-04-29
Knowledge cutoff2026-022025-082025-022024-10

Architecture & context

Context window1M400K200K131K
Max output tokens384K128K64K33K
Input modalitiestexttext, imagetext, imagetext
Reasoning modeoptionaloptionaloptionaloptional
Open weightsYesNoNoYes

Pricing (per Mtok)

Input$0.14$0.75$1$0.2
Output$0.28$4.5$5$0.6
Cached input$0.0028$0.075$0.1—
Batch input—$0.375$0.5—
Free tierYesNoYesYes

Speed

Output speed (tok/s)—180101.868
Time to first token (s)—0.50.860.78
Latency tierfastfastfastmedium

Trust & deployment

Trains on inputsYesNoNoNo
LicenseMITProprietaryProprietaryApache-2.0
Clouds—Azure OpenAI, Azure AI FoundryBedrock, Vertex AI, Azure AI FoundryGCP
Compliance—SOC2, GDPR, CCPA, HIPAASOC2, HIPAA, GDPR, ISO27001—

AI Panel scoring

Unified score8.798.48.5
Decision Maker8.598.58
Domain Strategist8.598.58.5
Finance Lead9.710109.5
Domain Practitioner8.598.59
Power User8887.5
Skeptic7.58.57.57
Value score9.99.89.89.5

Benchmarks

MMLU-Pro86.2—7682.8
GPQA Diamond88.1—67.270
AIME 2025———81.5
SWE-bench Verified79—73.3—
LiveCodeBench91.6——70.7
Terminal-Bench——41.75—
MRCR Long Context78.7———
SimpleQA34.1———
Humanity's Last Exam34.8———
LMArena Elo1438144814151423
LMArena Coding Elo1483149614821472

Scores link to their sources. Missing cells mean the vendor hasn't disclosed a result — honesty over padding.