UCP PlaygroundDeepSeek V3.2/Models/DeepSeek V3.2
InspectorAgentDocs
DeepSeekDeepSeek

DeepSeek V3.2

The highest checkout rate in the benchmark at 68%. Efficient at 6.2 turns average with low token usage. The trade-off: a 13.1% tool error rate, suggesting it brute-forces through errors rather than avoiding them. Never needs nudging — pushes through autonomously. The best raw completion rate of any model tested.

Avg Tokens69,072Avg Duration47.5sTurns to Checkout6.7
CheckoutCartSearchTurnsTokens22.137.176.46667
Shopping Score44/100
Fair

Scores computed from real agent sessions against live UCP-enabled stores. Not estimated — every data point is from an actual tool call, checkout attempt, and store response.

Shopping Score Breakdown

Checkout Rate40%22.1%
Cart Rate20%37.1%
Search Rate10%76.4%
Turn Efficiency15%66%
Token Efficiency15%67%
22.9% of sessions failed with errors.

Top Stores

StoreCheckout %Cart %
•••••••••••••••77.8%77.8%
houseofparfum.nl72%72%
monos.com50%50%
ucp.travel33.3%33.3%
allbirds.com20%20%
snocks.com0%0%
•••••••••••0%50%
•••••••••••••••••0%0%
••••••••••••••••••••••••••••••••0%100%
••••••••••••••••0%50%

Known Issues

Tool Error Ratemedium

13.1% error rate — brute-forces through errors rather than avoiding them.

JSON Formattinglow

Occasional issues with tool argument types. Requires extra validation overlays.

Token Usage

Avg per Session69,072
Fleet Average67,345
vs Fleet+3%
Median (p50)31,876
p90148,656
Prompt / Completion split
Prompt 68,312 (99%) Completion 760 (1%)
Daily Average Tokens (last 30 days)
Distribution range
Min2,517p2518,589p5031,876p7559,990p90148,656Max698,476

Test DeepSeek V3.2 on Your Store

Run a live agent session to see how DeepSeek V3.2 handles your store's checkout flow end-to-end.

Run Agent Session

Compare With Other Models

MetaMetaLlama 3.3 70B39.3%GoogleGoogleGemini 3.5 Flash38.8%AnthropicAnthropicClaude Sonnet 4.636.7%xAIxAIGrok 4.326.3%GoogleGoogleGemini 2.5 Flash25.7%AnthropicAnthropicClaude Opus 4.823.3%DeepSeekDeepSeekDeepSeek V3.222.1%GoogleGoogleGemini 3.1 Pro21%GoogleGoogleGemini 2.5 Pro20%OpenAIOpenAIGPT-4o14.8%OpenAIOpenAIGPT-5.511.1%xAIxAIGrok 3 Mini (Reasoning)11.1%OpenAIOpenAIo4-mini (Reasoning)9.4%AnthropicAnthropicClaude Sonnet 58.9%DeepSeekDeepSeekDeepSeek V4 Pro8.6%xAIxAIGrok 4.57.7%AnthropicAnthropicClaude Opus 56.8%DeepSeekDeepSeekDeepSeek V4 Flash6.5%Moonshot AIMoonshot AIKimi K34.5%DeepSeekDeepSeekDeepSeek R1 (Reasoning)4.1%GoogleGoogleGemini 3.6 Flash3.3%MetaMetaLlama 4 Maverick2.7%AlibabaAlibabaQwen 3.7 Max0%AlibabaAlibabaQwQ 32B (Reasoning)0%
Full comparison view →