DeepSeek
Frontier

DeepSeek-V3

Frontier

Open-weights frontier model that matched GPT-4-class quality at a fraction of the inference cost. Sparked the late-2024 cost reset.

DeepSeek-V3 is a frontier AI model from DeepSeek. It costs $0.2574 per million input tokens and $1.03 per million output tokens (blended $0.4888/M), with a 64,000-token context window.

RatesUSD per million tokens. Blend is 70% input, 30% output, our calculation.

INPUT
$0.2574/M
per million input tokens
OUTPUT
$1.03/M
per million output tokens
BLENDED 70/30
$0.4888/M

0.0%over 91 days · hover to read

DeepSeek-V3 — blended price

Reconstructed from 91 days of recorded rate-card changes. Prices hold flat between changes because that is what a posted price does — no value here is interpolated.

CONTEXT
64,000
tokens

From OpenRouter's listing, read on 17 Nov 2026.

What it is good at
  • Frontier quality, open weights
  • Aggressive pricing
  • Strong code & math
  • MoE architecture
Typical use cases
  • Self-hosted frontier inference
  • Cost-sensitive chat
  • Code generation

Benchmarks

vs. best public score
MMLU88.5%
Multitask academic knowledge across 57 subjects. Source: DeepSeek (self-reported), MMLU (EM), read 2026-11-10.
Graduate-level science questions, "Google-proof". Source: DeepSeek (self-reported), GPQA-Diamond (Pass@1), read 2026-11-10.
Real GitHub issues solved end-to-end. Source: DeepSeek (self-reported), SWE Verified (Resolved), read 2026-11-10.
LMArena Elo1358 Elo
Crowd-sourced head-to-head preference Elo rating. Source: LMArena (independent), LMArena Text Arena score, read 2026-11-10.
Hand-curated from each provider's published reports and public leaderboards. Methodology varies across sources, treat as directional rather than authoritative.

Also reported by the provider

Results on benchmarks outside the set we compare across models. They are listed for completeness and are deliberately excluded from the bars above and from every ranking on this site, because a score is only comparable against the same benchmark variant. Why benchmark scores do not compare →

  • MATH-500 (EM)90.2%self-reported by DeepSeek · 2024-12 · A 500-problem subset, not comparable with the full MATH benchmark, which DeepSeek reports only for the base model (61.6, 4-shot).
  • HumanEval-Mul (Pass@1)82.6%self-reported by DeepSeek · no date on the source · DeepSeek-V3 model card, chat-model table. Plain HumanEval is published only for the base model. Not comparable with the HumanEval column, so shown here and never ranked.

More from DeepSeek

See all 11 →

Price history

10 recorded changes
Recorded price changes for DeepSeek-V3, newest first
DateRateWasNowChange
14 Nov 2026Output$0.8900$1.03+16%
14 Nov 2026Input$0.3200$0.2574-20%
10 Nov 2026Output$1.03$0.8900-13%
10 Nov 2026Input$0.2574$0.3200+24%
31 Jul 2026Output$0.8001$1.03+29%
31 Jul 2026Input$0.2002$0.2574+29%
2 Jun 2026Output$0.9144$0.8001-12%
2 Jun 2026Input$0.2288$0.2002-13%
26 May 2026Input$0.3200$0.2288-29%
26 May 2026Output$0.8900$0.9144+2.7%

Rates per million tokens, as published by the provider or its router listing. Tracking runs since 3 May; anything earlier is not on record. See every tracked price change.

Frequently asked questions

How much does DeepSeek-V3 cost?

DeepSeek-V3 costs $0.2574 per million input tokens and $1.03 per million output tokens, for a blended reference rate of $0.4888 per million tokens.

What is DeepSeek-V3's context window?

DeepSeek-V3 supports up to 64,000 tokens of context in a single request.

What is DeepSeek-V3 best for?

DeepSeek-V3 is well suited to Frontier quality, open weights, Aggressive pricing and Strong code & math.

Who makes DeepSeek-V3?

DeepSeek-V3 is developed and served by DeepSeek.

Compared with

Terms used on this page