DeepSeek: DeepSeek V3

Public pricingIntelligence 79/100Medium memoryOutils

DeepSeek: DeepSeek V3 est un modèle texte conçu pour codage, débogage et travail technique. Il associe de bonnes performances en codage, un contexte de 164K tokens et un profil à coût équilibré pour un travail fiable sur coding, debugging, and technical writing.

Input

$0.32/1M

Output

$0.89/1M

Cached

$0.07/1M

Batch

$0.14/1M

Calculate your DeepSeek V3 bill.

Set your workload — see cost at your exact volume.

What would DeepSeek V3 cost you?

Adjust the workload to see your monthly bill.

1,00010,00050,000250,0001M10M

Technical specifications

DeepSeek V3 at a glance.

Memory

163,840

tokens

Max reply

163,840

tokens

Memory tier

Medium

a long report or a codebase file

Tokenizer

deepseek

Released

Dec 2024

Training cutoff

Jul 2024

Availability

Public pricing

Status

active

Benchmarks

Quality benchmarks

Independent evaluations from public leaderboards. Higher is better.

  • mmlu

    88.5

  • humaneval

    83

  • aa_intelligence_index

    16

  • gpqa_diamond

    67.61

  • scipredict

    19.18

What it can do

Capabilities & limits.

  • Understands images
  • Deep step-by-step thinking
  • Uses tools / calls functions
  • Strict JSON output
  • Streams replies
  • Fine-tunable on your data

When to pick DeepSeek V3

  • Agentic workflows that call tools or APIs.
  • High-volume workloads where unit cost matters.

When to look elsewhere

  • Your workload involves images — pick a vision-capable model instead.

Alternatives

Similar models to compare.

FAQ

DeepSeek V3 — the questions we see most.

Pricing, capabilities, alternatives — generated from the same data that powers the calculator above.

Get instant answers from our AI agent

At a typical workload of 50,000 conversations a month with 1,500-token prompts and 800-token replies, DeepSeek V3 costs roughly $60 per month. Input is $0.32 /1M tokens and output is $0.89 /1M tokens.
DeepSeek V3 has a 163,840-token context window (medium memory — a long report or a codebase file). That means you can fit about 30,720 words of input and history in a single call.
Beyond text generation, DeepSeek V3 supports calling functions / tools, strict JSON output. It streams replies by default.
DeepSeek V3 was released in December 2024, with training data cut off around July 2024.
Models in a similar class include DeepSeek V3.2 Exp, DeepSeek V3.2, DeepSeek V3.1 Terminus. The "Similar models" section below this FAQ links into each.

Still unsure?

Compare DeepSeek V3 against 100+ other models.

Open the full wizard — pick a use case, set your usage, and see side-by-side monthly costs in under a minute.

Open the comparison wizard