DeepSeek-V3.2 vs DeepSeek-V3.2-Speciale
Wins 2 of 6 areas
Agents · Long documents
Wins 3 of 6 areas
Coding · Reasoning · Following instructions
DeepSeek-V3.2-Speciale wins more areas, narrowly.DeepSeek-V3.2 is cheaper and better at agents.
Scores updated · 23 tests both models report · How we compare
Where each one wins
Tests won in each of the six 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.
- ReasoningHard problems that need careful thinking03DeepSeek-V3.2-Speciale3 of 3 tests
- CodingWriting and fixing software01DeepSeek-V3.2-Speciale1 of 3 tests · 2 ties
- Following instructionsDoing exactly what it is asked01DeepSeek-V3.2-Speciale1 of 1 test
- AgentsCarrying out multi-step tasks on its own10DeepSeek-V3.21 of 1 test
- Long documentsFinding answers in very long texts10DeepSeek-V3.21 of 1 test
- FactsGetting facts right instead of making them up11Even1 each
Agents, 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-V3.2 costs 3% less for the same work.
The biggest differences
The tests each model wins by the widest margin, up to three each. Scores are out of 100.
Where DeepSeek-V3.2-Speciale pulls ahead
- Code for real scientific research problemsSciCode+5.1points ahead
- Answers hard knowledge questions correctlyAA-Omniscience · Accuracy+4.8points ahead
- Research-level physics problemsCritPt+4.5points ahead
Where DeepSeek-V3.2 pulls ahead
- Avoids making up answers it doesn't knowAA-Omniscience · Non-hallucination+6.5points ahead
- Reasons across sets of long documentsAA-LCR+3.3points ahead
Every test, side by side
All 23 tests both models report. The winning score is in its model's colour; marks a score checked independently.
CodingDeepSeek-V3.2-Speciale
- SciCodeDeepSeek-V3.2-Speciale by 5.138.944+5.1
- Terminal-Bench Hardtie35.634.8tie
- SWE-bench Verifiedtie73.173.1tie
ReasoningDeepSeek-V3.2-Speciale
- CritPtDeepSeek-V3.2-Speciale by 4.52.97.4+4.5
- Humanity's Last ExamDeepSeek-V3.2-Speciale by 4.124.628.7+4.1
- GPQA DiamondDeepSeek-V3.2-Speciale by 3.18487.1+3.1
FactsEven
- AA-Omniscience · Non-hallucinationDeepSeek-V3.2 by 6.517.310.8+6.5
- AA-Omniscience · AccuracyDeepSeek-V3.2-Speciale by 4.83337.8+4.8
Following instructionsDeepSeek-V3.2-Speciale
- IFBenchDeepSeek-V3.2-Speciale by 3.260.763.9+3.2
Other results12 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.
- τ²-Bench Telecom (AA run)DeepSeek-V3.2 by 90.690.60+90.6
- AA Agentic IndexDeepSeek-V3.2 by 18.318.30+18.3
- HMMT 2025DeepSeek-V3.2-Speciale by 990.299.2+9
- AA IntelligenceDeepSeek-V3.2 by 721.514.5+7
- Artificial Analysis Coding IndexDeepSeek-V3.2 by 6.344.237.9+6.3
- HMMT Feb. 2025DeepSeek-V3.2-Speciale by 592.597.5+5
- AA-OmniscienceDeepSeek-V3.2-Speciale by 4.8-22.5-17.7+4.8
- HMMT Nov. 2025DeepSeek-V3.2-Speciale by 3.39093.3+3.3
- AIME 2025DeepSeek-V3.2-Speciale by 2.993.196+2.9
- t2-benchtie80.280.3tie
- Terminal-Bench 2.0tie46.446.4tie
- Toolathlontie35.235.2tie
Questions people ask
Which is better, DeepSeek-V3.2 or DeepSeek-V3.2-Speciale?
DeepSeek-V3.2-Speciale wins three of the six areas where both have results: coding, reasoning and following instructions. DeepSeek-V3.2 wins agents and long documents, and costs 3% less. They are level on facts.
Which is better for coding?
DeepSeek-V3.2-Speciale. It wins 1 of the 3 coding tests both models report; DeepSeek-V3.2 wins none, and 2 are ties.
Which is cheaper?
DeepSeek-V3.2 costs $0.28 per million input tokens and $0.42 per million output tokens; DeepSeek-V3.2-Speciale costs $0.29 and $0.43. That makes DeepSeek-V3.2 about 3% cheaper for the same work.
How do you compare the two?
We use the 23 benchmark tests both models have published scores on. The verdict counts the 11 tests in the eight capability areas, and a gap under one point (ten on rating-style scales) counts as a tie. The other 12 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.