AINews
CodingSkillsPromptsLearnCompareGlossaryBy roleTools
All tools
LLMs & chat AI

DeepSeek

By DeepSeek

Chinese open-weight model with strong reasoning and coding capabilities at very low cost.

Visit DeepSeekHas free tier

Overview

DeepSeek is the open-weights upstart that reshaped the price expectations of the entire LLM market. The flagship reasoning model is competitive with frontier closed models on math, coding, and general benchmarks — and the API costs roughly $0.27 per million input tokens and $1.10 per million output tokens. For comparison, Claude Opus is around $15/$75 per million. That's roughly 30x cheaper for the same task, with quality close enough that for many production workloads the math is obvious. DeepSeek's other big move is open weights. The models are downloadable and self-hostable, which means you can run them in your own VPC, on-prem, or on any cloud GPU you can get hold of — without sending data to a third party. For regulated industries, sovereign cloud needs, and privacy-sensitive workloads, this is a category by itself. The chat product at chat.deepseek.com is free, polished, and surprisingly fast. The tradeoffs are honest and worth knowing. Writing voice is more workmanlike than Claude or ChatGPT — fine for tasks, less great for publishable prose. The hosted service is operated from China, which is a non-starter for some enterprises and a reason others self-host the open weights instead. The third-party ecosystem (plugins, agents, marketplaces) is essentially nonexistent — you bring your own stack. But for raw cost-per-quality at scale, nothing currently competes.

Best for

  • cost-sensitive workloads
  • coding tasks
  • self-hosting

Strengths

  • ✓Roughly 30x cheaper than Claude or GPT-5 on the API for comparable task quality
  • ✓Open weights — self-host anywhere, no data leaves your environment
  • ✓Strong on math, coding, and structured reasoning benchmarks
  • ✓Free chat product is genuinely usable as a daily driver
  • ✓Available through every major inference provider for one-click hosted deployment

Weaknesses

  • ✗Writing voice is workmanlike — competent but not as polished as Claude or ChatGPT
  • ✗Hosted service runs from China, a hard no for some enterprise risk teams (mitigated by self-hosting)
  • ✗Essentially no first-party plugin/agent ecosystem
  • ✗Fewer multimodal features than frontier closed models (image and video understanding lag)

Pricing

Free chat

Free

Unlimited use of DeepSeek's flagship model at chat.deepseek.com. Genuinely free, with reasonable rate limits.

API — input

~$0.27 / 1M tokens

Frontier-tier reasoning model on the API. Roughly 30-50x cheaper than Claude or GPT-5 for input. Off-peak discounts further reduce cost.

API — output

~$1.10 / 1M tokens

Output tokens at a fraction of closed-model pricing. Long-form generation is dramatically more affordable.

Open weights (self-host)

Free + your compute

Download the model weights and run them anywhere — your VPC, on-prem, or any cloud GPU. Pay only for hardware. The strongest data-control option in AI.

Enterprise / partner deployments

Custom

Managed deployments through cloud partners (AWS Bedrock, Azure, Together, Fireworks, etc.) with SLAs, regional hosting, and compliance.

Use cases

  • High-volume API workloads where cost matters

    Anything that runs a frontier-tier model in a tight loop — bulk classification, summarization, code generation pipelines — drops in cost by an order of magnitude moving to DeepSeek.

  • Self-hosted AI for regulated or sensitive data

    Open weights mean you can run DeepSeek inside your own VPC with zero outbound data flow. The cleanest answer for healthcare, legal, defense, and sovereign cloud.

  • Coding assistant on a tight budget

    Strong coding performance at a fraction of Claude or GPT pricing — ideal for indie devs, students, and side projects where cost adds up fast.

  • Math and structured reasoning

    DeepSeek's reasoning model is genuinely competitive on hard math benchmarks. A solid choice for STEM tutoring tools and quantitative research.

  • Building startup MVPs without burning runway on inference

    If your product has frontier-quality LLM calls baked into the user flow, switching to DeepSeek can cut your cost-of-goods by 90%+ overnight.

  • Backup/fallback model in a multi-model architecture

    Many teams route the easy 80% of queries to DeepSeek and reserve Claude or GPT-5 for the hard 20%. Big cost savings, minimal quality hit.

When not to use

  • ✗You need the cleanest, most publishable writing voice — Claude is sharper
  • ✗Your enterprise risk team has flagged China-hosted services and you can't self-host the weights
  • ✗You want a rich plugin/agent ecosystem out of the box — ChatGPT or Gemini are far broader
  • ✗You need best-in-class image, video, or voice modalities — closed frontier models still lead here

Alternatives

Claude

A general-purpose chat assistant known for nuanced reasoning, careful writing, and very long context handling.

ChatGPT

The most widely-used AI chat assistant with image, voice, and a broad ecosystem of plugins and custom GPTs.

Gemini

Google's multimodal assistant with native integration into Gmail, Docs, Sheets, and the rest of Workspace.

Le Chat

European chat assistant from Mistral with strong multilingual support and a focus on data sovereignty.

Grok

xAI's assistant with real-time access to X (Twitter) data, useful for tracking live discourse.

See it compared

ChatGPT vs DeepSeekClaude vs DeepSeek

Glossary terms to know

LLM (Large Language Model)Frontier modelOpen weightsToken costContext window

Other LLMs & chat AI

Claude

A general-purpose chat assistant known for nuanced reasoning, careful writing, and very long context handling.

ChatGPT

The most widely-used AI chat assistant with image, voice, and a broad ecosystem of plugins and custom GPTs.

Gemini

Google's multimodal assistant with native integration into Gmail, Docs, Sheets, and the rest of Workspace.

Perplexity

A search-grounded answer engine that cites sources for every claim, useful when you need verifiable information.

Le Chat

European chat assistant from Mistral with strong multilingual support and a focus on data sovereignty.

Grok

xAI's assistant with real-time access to X (Twitter) data, useful for tracking live discourse.

Pi

Conversational AI tuned for warmth and reflection rather than raw output — designed to help you think through decisions, journal, or just talk something out.

Poe

Single interface to chat with most leading models (Claude, GPT, Gemini, Llama, etc.) plus thousands of community-built bots — pay one subscription, switch freely.