Model comparison

Claude Sonnet 4.6 vs DeepSeek-V3.2-Exp

Claude Sonnet 4.6 is the stronger model overall, scoring 50.3 to 44.3 on the Noometry Index. DeepSeek-V3.2-Exp costs 21× less per token, which makes it the better buy when Claude Sonnet 4.6's lead doesn't matter for your workload.

Last verified October 10, 2026. 39 shared benchmarks.

Claude Sonnet 4.6 Anthropic

50.3

Rank #50 Confirmed

DeepSeek-V3.2-Exp

DeepSeek

44.3

Rank #78 Confirmed

Summary

They share 39 benchmarks with published results for both. Claude Sonnet 4.6 scores higher in 7 categories and DeepSeek-V3.2-Exp in 2 categories; 7 gaps are clear of the uncertainty.
The widest gap is in reasoning, where Claude Sonnet 4.6 leads 46.1 to 22.1.
The biggest single-benchmark swing is ARC-AGI-2: 60.4% for Claude Sonnet 4.6 and 4% for DeepSeek-V3.2-Exp.
DeepSeek-V3.2-Exp is cheaper at $0.26 / $0.38 per million input/output tokens, against $3 / $15 for Claude Sonnet 4.6.
Claude Sonnet 4.6 accepts more context: 1M tokens versus 164K.
DeepSeek-V3.2-Exp has downloadable open weights; the other is API-only.

Side by side

Claude Sonnet 4.6 and DeepSeek-V3.2-Exp specifications
	Claude Sonnet 4.6	DeepSeek-V3.2-Exp
Provider	Anthropic	DeepSeek
Noometry Index	50.3	44.3
Released	2026-02-17	2025-09-29
Weights	Proprietary	Open
Context window	1M	164K
Max output	128K	66K
Input $ / M tokens	$3	$0.26
Output $ / M tokens	$15	$0.38
Results tracked	57	49

Sponsored placements are available on pages like this one. Advertise on Noometry

Category by category

Coding Too close to call

Claude Sonnet 4.6: 46.3 (#67), DeepSeek-V3.2-Exp: 46.5 (#65)

Coding benchmarks
Benchmark	Claude Sonnet 4.6	DeepSeek-V3.2-Exp
LMArena WebDev	1522	1362
SciCode	46.8%	38.9%
WeirdML	66.1%	39.5%
LMArena Coding	1504	1454
SWE-bench Verified	75.2%	—
DeepSWE	29.9%	—
FrontierCode	24.3%	—
SWE-bench Verified (bash only)	—	70%
Aider Polyglot	—	74.2%
SWE-bench Multilingual	—	59%
ALE-Bench	1,327	—

Agentic & Tool Use Claude Sonnet 4.6 leads

Claude Sonnet 4.6: 39.1 (#28), DeepSeek-V3.2-Exp: 32.7 (#59)

Agentic & Tool Use benchmarks
Benchmark	Claude Sonnet 4.6	DeepSeek-V3.2-Exp
Terminal-Bench	53.4%	39.6%
APEX-Agents	43%	21.3%
Vending-Bench 2	7,204	1,034
Berkeley Function Calling Leaderboard	—	56.7%
OSWorld 2.0	9.3%	—
TheAgentCompany	—	42.9%
DeepResearch Bench	54.9%	—
OSWorld	72.1%	—
ExploitBench	23.6%	—
GBAEval	48.8%	—
GDP.pdf	18%	—
LMArena Search	1221	—

Reasoning Claude Sonnet 4.6 leads

Claude Sonnet 4.6: 46.1 (#45), DeepSeek-V3.2-Exp: 22.1 (#208)

Reasoning benchmarks
Benchmark	Claude Sonnet 4.6	DeepSeek-V3.2-Exp
ARC-AGI-2	60.4%	4%
NYT Connections (extended)	80.9%	36.7%
ARC-AGI-1	86.5%	57%
CritPt	3.1%	2.9%
Chess Puzzles	13%	14%
Thematic Generalization	76.3%	65%
LMArena Hard Prompts	1484	1434
DTBench	89.9%	87.7%
LMCA	46.5%	29.1%
Epoch Capabilities Index	152.24	146.27
Kagi LLM Benchmark	—	52.2%
Mystery Game Puzzles	16%	—
ForecastBench	62	—

Math Claude Sonnet 4.6 leads

Claude Sonnet 4.6: 52.9 (#49), DeepSeek-V3.2-Exp: 41.7 (#87)

Math benchmarks
Benchmark	Claude Sonnet 4.6	DeepSeek-V3.2-Exp
OTIS Mock AIME 2024-2025	85.8%	87.8%
ProofBench	45%	8%
LMArena Math	1462	1435
FrontierMath (Feb 2025 set)	32.4%	22.1%
FrontierMath Tier 4 (v1)	8.3%	2.1%
MathArena Final-Answer Competitions	—	57.7%

Knowledge Too close to call

Claude Sonnet 4.6: 51.7 (#65), DeepSeek-V3.2-Exp: 51.7 (#66)

Knowledge benchmarks
Benchmark	Claude Sonnet 4.6	DeepSeek-V3.2-Exp
GPQA Diamond	87.4%	83.4%
Vectara Hallucination Rate	10.6%	5.3%
LMArena Expert	1500	1436
SimpleQA Verified	35.5%	—

Multimodal Not comparable

Claude Sonnet 4.6: 38.0 (#68), DeepSeek-V3.2-Exp: —

Multimodal benchmarks
Benchmark	Claude Sonnet 4.6	DeepSeek-V3.2-Exp
LMArena Vision	1283	—
Blueprint-Bench 2	6.7%	—
LMArena Document	1482	—

Multilingual Claude Sonnet 4.6 leads

Claude Sonnet 4.6: 54.4 (#41), DeepSeek-V3.2-Exp: 52.2 (#90)

Multilingual benchmarks
Benchmark	Claude Sonnet 4.6	DeepSeek-V3.2-Exp
LMArena Non-English	1440	1409
LMArena Chinese	1491	1461
LMArena French	1465	1433
LMArena German	1428	1440
LMArena Japanese	1420	1374
LMArena Korean	1411	1371
LMArena Russian	1440	1424
LMArena Spanish	1464	1440

Instruction Following Claude Sonnet 4.6 leads

Claude Sonnet 4.6: 77.4 (#25), DeepSeek-V3.2-Exp: 74.5 (#93)

Instruction Following benchmarks
Benchmark	Claude Sonnet 4.6	DeepSeek-V3.2-Exp
LMArena Instruction Following	1475	1413

Long Context DeepSeek-V3.2-Exp leads

Claude Sonnet 4.6: 45.3 (#44), DeepSeek-V3.2-Exp: 47.6 (#16)

Long Context benchmarks
Benchmark	Claude Sonnet 4.6	DeepSeek-V3.2-Exp
LMArena Longer Query	1479	1428
Fiction.LiveBench	—	83.3%
CL-bench	—	13.2%
CL-bench Life	—	9.5%

Writing & Preference Claude Sonnet 4.6 leads

Claude Sonnet 4.6: 70.2 (#22), DeepSeek-V3.2-Exp: 62.4 (#77)

Writing & Preference benchmarks
Benchmark	Claude Sonnet 4.6	DeepSeek-V3.2-Exp
LMArena Text	1458	1425
LMArena Creative Writing	1435	1403
EQ-Bench Creative Writing	1810	1515
LMArena Multi-Turn	1464	1427
EQ-Bench 4	1207	—

Frequently asked questions

Is Claude Sonnet 4.6 better than DeepSeek-V3.2-Exp?

Which is cheaper, Claude Sonnet 4.6 or DeepSeek-V3.2-Exp?

DeepSeek-V3.2-Exp is cheaper. It lists at $0.26 per million input tokens and $0.38 per million output tokens; Claude Sonnet 4.6 lists at $3 and $15.

Is Claude Sonnet 4.6 or DeepSeek-V3.2-Exp better for coding?

They score almost the same on coding (46.3 vs 46.5); test both on your own repository before choosing.

Which has the bigger context window?

Claude Sonnet 4.6 does, with 1M tokens against 164K.

How many benchmarks do Claude Sonnet 4.6 and DeepSeek-V3.2-Exp share?

39 benchmarks have published results for both models. Claude Sonnet 4.6 has 57 scored results on Noometry and DeepSeek-V3.2-Exp has 49.