Which is better, DeepSeek-V4-Flash or GPT-5.4?

Across 50 shared benchmarks tracked on Vector Wire, DeepSeek-V4-Flash leads on 12 and GPT-5.4 leads on 38 — GPT-5.4 leads on more. Strengths differ by axis; the table below shows every head-to-head number with provenance.

SEEDRefresh Models —Deals —Regulatory —Sources —

DeepSeek-V4-Flash vs GPT-5.4

DeepSeekvsOpenAI50 shared benchmarks12–38 head-to-head

Benchmark	DeepSeek-V4-Flash	GPT-5.4
AA Agentic Index	61.3	58.2
AA Intelligence	51.8	53.1
AA-LCR	74.3	77.7
AA-Omniscience	-14.3	5.8
aime_2026	95.8	99.2
Apex (Pass@1)	33	54.1
Apex Shortlist (Pass@1)	85.7	78.1
arc_agi_1	84	93.7
ARC-AGI-2	46	74
arena_elo	1439	1477
arena_text_factuality	1434	1476
Artificial Analysis Coding Index	69.1	71.1
browsecomp	73.2	82.7
Chinese SimpleQA (C-SimpleQA)	78.9	76.8
Codeforces	3052	3168
coding_arena_elo	1431	1457
critpt	16.6	23.4
cybergym	76.7	66.3
gdpval	52.9	50.1
GDPval-AA (Elo)	1395	1674
GPQA Diamond	90.8	93
HLE	38.6	43.7
HLE (with tools)	45.1	52.1
hmmt_feb_2026	94.8	97.7
IFBench	79.2	73.9
imo_answer_bench	88.4	91.4
itbenchSre	31.5	34.5
livebench_agentic_coding	46.8	53.8
livebench_coding	75	77.5
livebench_data_analysis	79.3	79.3
livebench_instruction_following	65.5	70.2
livebench_language	79.2	82.6
livebench_math	86.8	94.2
livebench_reasoning	86.6	88.1
MCP Atlas	69	70.6
MMLU-Pro	86.4	87.5
nl2repo	54.2	41.3
OmniScience Accuracy	40.4	50.9
OmniScience Non-Hallucination	15.6	17.4
scicode	49.9	56.6
simpleqa_verified	45	45.3
SWE-bench Multilingual	73.3	71.7
SWE-bench Pro	52.6	59.1
TauBench V3 - Banking	39.4	39.6
Terminal-Bench 2.0	56.9	75.1
Terminal-Bench 2.1	82.7	78.3
Terminal-Bench Hard	38.6	57.6
toolathlon	70.3	54.6
τ²-Bench Telecom (AA run)	95.6	87.1
τ³-Bench	22.9	72.9

Best tracked score per model per benchmark (default configuration; source-attributed). ↓ marks lower-is-better metrics. Open either model for its full surface, provenance and pricing.