DeepSeek-V4-Flash vs Qwen3.5 397B A17B
Wins 6 of 7 areas
Coding · Agents · Reasoning · Facts · Math · Long documents
Wins 0 of 7 areas
—
DeepSeek-V4-Flash is the stronger all-rounder.
Scores updated · 54 tests both models report · How we compare
Where each one wins
Tests won in each of the seven areas where both have results. Each piece is one test, so a longer bar means more evidence; grey means the two scored within a point of each other.
- CodingWriting and fixing software71DeepSeek-V4-Flash7 of 8 tests
- MathCompetition and research-level math40DeepSeek-V4-Flash4 of 4 tests
- AgentsCarrying out multi-step tasks on its own41DeepSeek-V4-Flash4 of 5 tests
- ReasoningHard problems that need careful thinking30DeepSeek-V4-Flash3 of 3 tests
- FactsGetting facts right instead of making them up21DeepSeek-V4-Flash2 of 3 tests
- Long documentsFinding answers in very long texts10DeepSeek-V4-Flash1 of 1 test
- Following instructionsDoing exactly what it is asked00Even0 each · 1 tie
Long documents and following instructions rest on a single test each.
What it costs
Prices per million tokens, roughly 750,000 words. The bars show the cost of a million tokens read plus a million written.
DeepSeek-V4-Flash costs 58% less for the same work.
The biggest differences
The three tests each model wins by the widest margin. Scores are out of 100.
Where DeepSeek-V4-Flash pulls ahead
- Real work tasks from 44 professionsGDPVal+32.1points ahead
- Command-line tasks in a real terminalTerminal-Bench 2.1+27.4points ahead
- Unpublished advanced math problemsFrontierMath Tiers 1-3 (v2)+26.3points ahead
Where Qwen3.5 397B A17B pulls ahead
- Hard command-line tasks in a real terminalTerminal-Bench Hard+5.3points ahead
- Avoids making up answers it doesn't knowAA-Omniscience · Non-hallucination+2.8points ahead
- Finds the root cause of IT system incidentsAA IT-Bench SRE+2.6points ahead
Every test, side by side
All 54 tests both models report. The winning score is in its model's colour; marks a score checked independently.
CodingDeepSeek-V4-Flash
- Terminal-Bench 2.1DeepSeek-V4-Flash by 27.478.751.3+27.4
- LiveCodeBench v6DeepSeek-V4-Flash by 790.683.6+7
- SciCodeDeepSeek-V4-Flash by 5.550.344.8+5.5
- Terminal-Bench HardQwen3.5 397B A17B by 5.335.640.9+5.3
- SWE-bench MultilingualDeepSeek-V4-Flash by 473.369.3+4
- LMArena · WebDevDeepSeek-V4-Flash by 30 rating points14301400+30 rating
- SWE-bench VerifiedDeepSeek-V4-Flash by 2.67976.4+2.6
- SWE-bench ProDeepSeek-V4-Flash by 1.752.650.9+1.7
AgentsDeepSeek-V4-Flash
- GDPValDeepSeek-V4-Flash by 32.146.914.8+32.1
- τ-Bench V3 · BankingDeepSeek-V4-Flash by 2639.413.4+26
- Terminal-Bench 4.0DeepSeek-V4-Flash by 12.112.10+12.1
- BrowseCompDeepSeek-V4-Flash by 4.273.269+4.2
- AA IT-Bench SREQwen3.5 397B A17B by 2.631.534.1+2.6
ReasoningDeepSeek-V4-Flash
- CritPtDeepSeek-V4-Flash by 14.916.61.7+14.9
- Humanity's Last ExamDeepSeek-V4-Flash by 9.638.629+9.6
- GPQA DiamondDeepSeek-V4-Flash by 1.590.889.3+1.5
FactsDeepSeek-V4-Flash
- AA-Omniscience · AccuracyDeepSeek-V4-Flash by 9.640.430.8+9.6
- SimpleQA VerifiedDeepSeek-V4-Flash by 8.134.126+8.1
- AA-Omniscience · Non-hallucinationQwen3.5 397B A17B by 2.88.311.1+2.8
MathDeepSeek-V4-Flash
- FrontierMath Tiers 1-3 (v2)DeepSeek-V4-Flash by 26.357.531.2+26.3
- IMOAnswerBenchDeepSeek-V4-Flash by 7.588.480.9+7.5
- HMMT Feb. 2026DeepSeek-V4-Flash by 6.994.887.9+6.9
- AIME 2026DeepSeek-V4-Flash by 4.595.891.3+4.5
Long documentsDeepSeek-V4-Flash
- AA-LCRDeepSeek-V4-Flash by 2.479.777.3+2.4
Other results29 tests, not counted
Tests outside the eight areas. They are not counted above: several are summary scores built from other tests, or the same test under another name.
- ToolathlonDeepSeek-V4-Flash by 3270.338.3+32
- AA Agentic IndexDeepSeek-V4-Flash by 31.141.710.6+31.1
- Toolathlon VerifiedDeepSeek-V4-Flash by 29.670.340.7+29.6
- GDPval-AA v2DeepSeek-V4-Flash by 227 rating points1189962+227 rating
- NL2RepoDeepSeek-V4-Flash by 2254.232.2+22
- Apex-Shortlist (with tools)DeepSeek-V4-Flash by 21.68260.4+21.6
- Apex-Shortlist (no tools)DeepSeek-V4-Flash by 2182.461.4+21
- Artificial Analysis Coding IndexDeepSeek-V4-Flash by 20.969.148.2+20.9
- AA-OmniscienceDeepSeek-V4-Flash by 16.4-14.3-30.8+16.4
- AA IntelligenceDeepSeek-V4-Flash by 15.934.318.4+15.9
- τ³-Bench BankingDeepSeek-V4-Flash by 9.522.913.4+9.5
- CritPt (no tools)DeepSeek-V4-Flash by 8.210.62.4+8.2
- BrowseComp (context management)Qwen3.5 397B A17B by 5.473.278.6+5.4
- IMOAnswerBench (with tools)DeepSeek-V4-Flash by 5.189.684.5+5.1
- PinchBenchDeepSeek-V4-Flash by 4.791.386.6+4.7
- Terminal-Bench 2.0DeepSeek-V4-Flash by 4.456.952.5+4.4
- TauBench V3 - AirlineDeepSeek-V4-Flash by 4.380.876.5+4.3
- ProfBench (Search)DeepSeek-V4-Flash by 45753+4
- HLE (with tools)Qwen3.5 397B A17B by 3.245.148.3+3.2
- Vals.ai Financial Agent 1.1 - without web searchQwen3.5 397B A17B by 2.958.461.3+2.9
- TauBench V3 - AverageDeepSeek-V4-Flash by 2.773.771+2.7
- StrongREJECTQwen3.5 397B A17B by 297.499.4+2
- SWEBench Pro PublicDeepSeek-V4-Flash by 1.752.650.9+1.7
- MMLU-ProQwen3.5 397B A17B by 1.686.287.8+1.6
- GPQA (unspecified)DeepSeek-V4-Flash by 1.488.587.1+1.4
- Vals.ai Financial Agent 1.1 - with web searchDeepSeek-V4-Flash by 1.160.159+1.1
- TauBench V3 - Retailtie89.188.5tie
- τ²-Bench Telecom (AA run)tie9595.6tie
- SciCode (subtask)tie48.248tie
Questions people ask
Which is better, DeepSeek-V4-Flash or Qwen3.5 397B A17B?
DeepSeek-V4-Flash wins six of the seven areas where both have results: coding, agents, reasoning, facts, math and long documents. Qwen3.5 397B A17B wins none. They are level on following instructions.
Which is better for coding?
DeepSeek-V4-Flash. It wins 7 of the 8 coding tests both models report; Qwen3.5 397B A17B wins 1.
Which is cheaper?
DeepSeek-V4-Flash costs $0.44 per million input tokens and $1.32 per million output tokens; Qwen3.5 397B A17B costs $0.60 and $3.60. That makes DeepSeek-V4-Flash about 58% cheaper for the same work.
How do you compare the two?
We use the 54 benchmark tests both models have published scores on. The verdict counts the 25 tests in the eight capability areas, and a gap under one point (ten on rating-style scales) counts as a tie. The other 29 are listed but not counted, because several are summary scores or repeat a test. Each score is the one shown on the model's own page.