DeepSeek V4 Pro
DeepSeekOpen weightsMITReleased Aug 13, 2026
On the arena.ai Agents leaderboard, DeepSeek V4 Pro has a task success rate of 0.36% (High, rank #30, 95% interval -1.41%–2.13%). DeepSeek V4 Pro ranks #56 on the LMArena Text leaderboard (style control) with an Arena score of 1464.6 (±6.6, 9,800 votes). Artificial Analysis Intelligence Index: 36.0. API price: $1.32 input / $3.96 output per 1M tokens. Context: 1.05M. Nearest rivals: GLM-5.1 (1465.2) above and Claude Sonnet 5 (1461.1) below.
Quick answer
- Agents leaderboard
- #30 of 48 · 0.36% success
- Text (LMArena)
- #56 · 1464.6
- 17NAS overall
- #30 of 84
- API price (in / out)
- $1.32 / $3.96 per 1M tokens
- Closest rival
- GLM-5.1 · 1465.2
- Cheaper alternative
- MiMo-V2.6-Pro: Text 1479.2, output $0.87 (78% cheaper)
- Data updated
- Oct 16, 2026 · synced daily
Sources: arena.ai, LMArena, Artificial Analysis, OpenRouter
Key metrics
Reasoning effort comparison
Artificial Analysis measurements per reasoning effort of the same model| Effort | Intelligence | Coding | tok/s | TTFT (s) | First answer (s) |
|---|---|---|---|---|---|
| maxUsed in table | 36.0 | 68.8 | 93 | 1.2 | 23 |
| non-reasoning | 20.4 | – | 94 | 1.1 | 1.1 |
Models ranked nearby
Sorted by Arena score| # | Model | Arena | Intel. | Output $ | tok/s |
|---|---|---|---|---|---|
| 40 |  GPT-5.6 Terra | 1466.4 | 42.1 | $12.00 | 116 |
| 41 |  Grok 4.5 | 1465.7 | 38.8 | $6.00 | – |
| 42 |  GLM-5.1 | 1465.2 | 26.1 | $4.07 | – |
| 43 | DeepSeek V4 ProThis page | 1464.6 | 36.0 | $3.96 | 93 |
| 44 |  Claude Sonnet 5 | 1461.1 | 38.2 | $10.00 | – |
| 45 |  Kimi K2.6 | 1460.8 | 27.0 | $4.00 | – |
| 46 |  Qwen3.6 Max | 1460.6 | 28.4 | – | – |
More from DeepSeek
| # | Model | Arena | Intel. | Output $ | tok/s |
|---|---|---|---|---|---|
| 32 | DeepSeek V4.1 Flash | 1474.6 | 39.5 | $1.20 | 180 |
| 61 | DeepSeek V4 Flash | 1438.7 | 34.3 | $1.32 | – |
| 67 | DeepSeek V3.2 | 1424.6 | 16.0 | $0.42 | – |
| – | DeepSeek V4 Flash Vision | – | 34.8 | $1.32 | 176 |
| – | DeepSeek V4 Pro 0424 | – | 30.1 | $0.87 | – |
Pricing
- Official API list price: $1.32 input and $3.96 output per 1M tokens (USD).
- At a typical 3:1 input-to-output ratio, the blended cost is about $1.98 per 1M tokens.
- Among 79 models with an output price, it is the #31 cheapest.
- The cheapest OpenRouter route charges $0.96 input / $1.91 output.
Sources & methodology
On the arena.ai Agents leaderboard, DeepSeek V4 Pro has a task success rate of 0.36% (High, rank #30, 95% interval -1.41%–2.13%).
- arena.ai Agents — agent task success rate, 95% interval, sessions and rank (via the public Hugging Face dataset, CC BY 4.0).
- LMArena — Arena score (Text leaderboard with style control, the same view arena.ai shows by default), 95% confidence interval, votes and rank.
- Artificial Analysis — Intelligence / Coding Index, output speed and latency (medians).
- OpenRouter — Route prices, context length and listing date.
Arena scores use the Bradley-Terry model (not Elo); ± is the half-width of the 95% confidence interval. Latency figures are Artificial Analysis medians; for reasoning models, time to first token and to first answer include the thinking phase. Data syncs daily at 14:20 (UTC+8).
Limits: Arena scores come from anonymous users voting on chats, so they reflect conversational preference rather than agent or coding benchmarks; with few votes the interval is wide and ranks move. Prices and speeds change as providers update them.
FAQ
How does DeepSeek V4 Pro do on the Agents leaderboard?
On the arena.ai Agents leaderboard, DeepSeek V4 Pro has a task success rate of 0.36% (High, rank #30, 95% interval -1.41%–2.13%).
What rank is DeepSeek V4 Pro right now?
DeepSeek V4 Pro is #43 on our leaderboard of 69 scored models (Arena score 1464.6); its raw LMArena rank is #56.
What does DeepSeek V4 Pro cost via API?
Input $1.32 / output $3.96 per 1M tokens (official API list price; the OpenRouter quote when no official price is available). At a typical 3:1 input-to-output ratio, the blended cost is about $1.98 per 1M tokens.
How fast is DeepSeek V4 Pro?
Artificial Analysis measures about 93 output tokens per second, 1.2 s to the first token and 23 s to the first answer token; it is a reasoning model, so latency includes thinking time.
How close is DeepSeek V4 Pro to the models ranked next to it?
It is 0.6 points behind GLM-5.1 (1465.2) and 3.5 points ahead of Claude Sonnet 5 (1461.1); gaps smaller than the confidence interval (±6.6) are not significant.
Is DeepSeek V4 Pro open source? Can it be self-hosted?
DeepSeek V4 Pro is an open-weights model (license: MIT); you can download the weights and self-host.
What is the context length of DeepSeek V4 Pro?
1.05M (1,048,576 tokens, as listed on OpenRouter).
DeepSeek V4 Pro vs GLM-5.1: which is better?
On the LMArena Text leaderboard DeepSeek V4 Pro scores 1464.6 and GLM-5.1 scores 1465.2 (a 0.6-point gap, within the confidence interval, so not significant). On price, DeepSeek V4 Pro costs $3.96 and GLM-5.1 $4.07 per 1M output tokens.
Is there a cheaper model close to DeepSeek V4 Pro?
Yes: MiMo-V2.6-Pro scores 1479.2 on the Text leaderboard (DeepSeek V4 Pro: 1464.6) and costs $0.87 per 1M output tokens, about 78% less than DeepSeek V4 Pro ($3.96).
Where does the DeepSeek V4 Pro data come from and how often is it updated?
Agent success rates come from the arena.ai Agents leaderboard (CC BY 4.0), Text scores from LMArena, intelligence and speed from Artificial Analysis, and prices and context from official list prices and OpenRouter. 17NAS syncs daily at 06:20 UTC; this page was last updated Oct 16, 2026.