DeepSeek V4 Pro benchmarks

DeepSeek V4 Pro is an open-weights model from DeepSeek, released 13 Aug 2026. It ranks 38th of 51 on our overall leaderboard and 20th of 41 for agents. At A$2.78 per million tokens (blended), it costs about 1.3× cheaper than the median model we track. It writes about 79 tokens a second, and answers start after 27 s on average, thinking included. It can be called from Microsoft Azure in Australia East, but only through global routing, so requests may be processed outside Australia; its weights are public, so it can also be self-hosted in Australia.

#38 OF 51 OVERALLA$2.78 / 1M TOKENS1M CONTEXTUPDATED 23 SEP 2026
CONTEXT WINDOW
1,000,000 tokens
MAX OUTPUT
384,000 tokens
INPUT
text
WEIGHTS
Open (MIT)
RELEASED
13 Aug 2026
OUTPUT SPEED
79 tok/s
ANSWER STARTS AFTER
27 s (thinking included)
DEEPSEEK DOCS
§ 02 — RESULTS

Benchmark results.

Independent test scores are for the reasoning, max effort setting. “Among tracked” ranks DeepSeek V4 Pro against the 51 models on this site; LMArena ranks run across its full leaderboard.

ALL RESULTS8 OF 8 CORE SIGNALS
DeepSeek V4 Pro benchmark results
BENCHMARKRESULTAMONG TRACKEDSOURCE DETAIL
INDEPENDENT TESTS · ARTIFICIAL ANALYSIS
AA Intelligence Index36.0#31 of 51Reasoning, max effort
AA Coding Index68.8#28 of 42Reasoning, max effort
GPQA Diamond92.8%#14 of 44Reasoning, max effort
Humanity’s Last Exam41.0%#30 of 51Reasoning, max effort
AA-LCR80.3%#24 of 51Reasoning, max effort
Terminal-Bench 2.178.7%#24 of 42Reasoning, max effort
τ²-Bench (banking)39.6%#16 of 41Reasoning, max effort
SciCode51.0%#35 of 46Reasoning, max effort
BLIND HUMAN VOTES · LMARENA
LMArena Text9,008 votes · ±71463#31 of 42#50 of 402 on LMArena · 13 Sep 2026
LMArena WebDev4,636 votes · ±101580#19 of 45#23 of 129 on LMArena · 22 Sep 2026
LMArena Agent47,602 sessions#16#14 of 37#16 of 46 on LMArena · 15 Sep 2026
§ 03 — COST

What DeepSeek V4 Pro costs in Australian dollars.

List prices converted at the snapshot’s RBA rate, excluding GST. Reasoning models are billed for their thinking as output tokens, so heavier settings cost more than the job estimates show.

LIST PRICE
DeepSeek V4 Pro price per million tokens
PER MILLION TOKENSAUDUSD LIST
Input tokensA$1.85US$1.32
Output tokensA$5.56US$3.96
Blended (3 in : 1 out)A$2.78US$1.98
EVERYDAY JOBSESTIMATES
Estimated cost of common jobs in Australian dollars
JOBCOST (AUD)MEDIAN MODEL
Customer support replyPER 1,000 REPLIESA$5.37A$6.58
Summarise a 30-page documentPER 100 DOCUMENTSA$4.15A$4.74
Agentic coding taskPER 10 TASKSA$3.22A$3.72
§ 04 — AUSTRALIA

Running DeepSeek V4 Pro in Australia.

It can be called from Microsoft Azure in Australia East, but only through global routing, so requests may be processed outside Australia; its weights are public, so it can also be self-hosted in Australia. How the platforms compare →

AUSTRALIAN CLOUD REGIONS
DeepSeek V4 Pro availability in Australian cloud regions
PLATFORMAVAILABILITYDETAIL
AWS Bedrock · SydneyNOT OFFERED
AWS Bedrock · MelbourneNOT OFFERED
Azure · Australia EastGLOBAL

Provider docs, checked 23 Sep 2026

Google Vertex AI · SydneyNOT OFFERED
Your own infrastructureSELF-HOST

Open weights: run it on your own servers or GPU instances in an Australian region.

§ 06 — QUESTIONS

DeepSeek V4 Pro, answered.

DeepSeek V4 Pro lists at US$1.32 per million input tokens and US$3.96 per million output tokens — A$1.85 and A$5.56 at A$1 = US$0.7123 (RBA, 22 Sep 2026), excluding GST. A typical customer support reply works out at about A$5.37 per 1,000 replies, before any reasoning tokens.

DeepSeek V4 Pro can be called from Microsoft Azure in Australia East, but only through global routing, so requests may be processed outside Australia; its weights are public, so it can also be self-hosted in Australia. Availability comes from the cloud providers’ own documentation; check your provider’s current terms before relying on it for data-residency obligations.

It ranks 25th of 42 on our coding ranking, with an Artificial Analysis Coding Index of 68.8 and an LMArena WebDev rating of 1580 (23rd of 129). The current leader is Claude Fable 5.1.

It ranks 20th of 41 on our agents ranking, completing 39.6% of τ²-Bench banking customer-service tasks and 78.7% of Terminal-Bench 2.1 tasks.

Artificial Analysis measures DeepSeek V4 Pro at about 79 tok/s of output, with the answer starting after 27 s on average once thinking time is included (reasoning, max effort setting).

DeepSeek documents a context window of 1,000,000 tokens (1M), with up to 384,000 output tokens. On AA-LCR, which tests reasoning across ~100,000-token document sets, it scores 80.3%.

PUT THE COMPARISON TO WORK

Need help choosing and using AI for your business?