DeepSeek V3.2 vs DeepSeek V4 Pro 0813

Wins 1 of 4 benchmarks
$0.36 / task
Wins 1 of 4 benchmarks
Humanity's Last Exam Humanity's Last Exam: Very hard expert questions across many fields, written to stump today’s models
DeepSeek V3.211.2
DeepSeek V4 Pro 081310.6
Terminal-Bench 4.0 Terminal-Bench 4.0 · source setting 2: Agents use a terminal to complete executable coding, system administration and research tasks in the 4.0 suite
DeepSeek V3.2—
DeepSeek V4 Pro 08131.5
AutomationBench-AA AutomationBench-AA: Agents orchestrate SaaS workflows through APIs; objectives receive credit only when the task has no guardrail violation
DeepSeek V3.2—
DeepSeek V4 Pro 081334.9
Arena Text · creative writing Arena Text · creative writing: Human pairwise preferences in the stated text vote pool
DeepSeek V3.21,400
DeepSeek V4 Pro 0813 · high1,446

Specs

SpecDeepSeek V3.2DeepSeek V4 Pro 0813
Input, per 1M$0.28!$1.32
Output, per 1M$0.40!$3.96
Cost per task?—$0.36
Speed, tok/s—137
Context128K1M
WeightsOpenOpen
ReleasedDec 2025Aug 2026
Memory (4-bit)411 GB990 GB
LicenceMITMIT
Commercial useAllowedAllowed
See images——
Read PDFsPartial—
Hear audio——
Watch video——
Call tools✓✓
Force a tool callPartial✓
Return strict JSON✓✓
Search the web??
Run code??
Use a computer??

What breaks when you switch

DeepSeek V3.2 → DeepSeek V4 Pro 0813
Memory (4-bit): 411 GB → 990 GB
Pass back reasoning_content · Chat Completions›

With thinking and tools, return all prior reasoning_content even on turns without tool calls or the API returns HTTP 400.

Disable thinking mode first · Chat Completions›

Required and named tool choices return HTTP 400 in thinking mode; disable thinking first.

Parameters have no effect›

In thinking mode, temperature, presence_penalty and frequency_penalty have no effect.

top_p value is ignored›

In thinking mode, top_p is fixed at 0.95 with tools and 1.0 otherwise; in non-thinking mode it takes effect.

Stateless API · Responses›

Responses does not support previous_response_id or conversation; preserve conversation history in your client.

DeepSeek V4 Pro 0813 → DeepSeek V3.2
Context: 1M → 128K