agenticoutputs
Frontier AI, Open Weights

DeepSeek: the open-weight lab that made frontier AI 10x cheaper and MIT-licensed.

In January 2025, DeepSeek-R1 matched OpenAI o1 on reasoning benchmarks at 1/20th the training cost and triggered what Marc Andreessen called "AI's Sputnik moment." By July 2026, V4 Pro ships at 34x less than GPT-5.5 per token - with 1M context, open weights, and an MIT license. This is the full picture.

V3 Nov 2024 671B MoE R1 Jan 2025 Sputnik moment V3.2 Mar 2026 128K context V4 Flash Apr 2026 284B MoE · 1M ctx $0.14/$0.28 V4 Pro Apr 2026 1.6T MoE · 1M ctx $0.44/$0.87

Five model generations in 18 months. The V4 launch on April 24, 2026 consolidated the entire lineup into two models: Flash for throughput and Pro for reasoning. Legacy aliases (deepseek-chat, deepseek-reasoner) retire July 24, 2026.

01 What It Is

A Chinese AI research lab that ships open-weight models under MIT license and prices them to move.

DeepSeek was founded in 2023 by Liang Wenfeng, spun out of the AI hedge fund High-Flyer. The first public model, DeepSeek Coder, shipped in November 2023. By January 2025, DeepSeek-R1 had matched OpenAI o1 on reasoning benchmarks - trained for an estimated $5.6M in compute, roughly 1/20th of what comparable US models cost.

Three things set DeepSeek apart. First, every model from V3 forward is MIT licensed: download the weights, fine-tune them, host them yourself, build commercial products. Second, Mixture-of-Experts architecture activates only a fraction of total parameters per token (V4 Pro: 49B of 1.6T; V4 Flash: 13B of 284B), keeping inference cheap. Third, API pricing consistently undercuts US providers by 10-34x.

02 The Models

Two V4 models replaced the entire lineup.

Everything consolidated into V4 Flash and V4 Pro on April 24, 2026. R1 and V3.2 are historical references. The legacy API aliases retire July 24, 2026.

V4 Flash

The Workhorse

Params284B total / 13B active
Context1M tokens
Max Output384K tokens
Price$0.14 / $0.28 per 1M
Cache Hit$0.0028 per 1M

Default for production workloads: chat, RAG, extraction, classification, translation, tool-calling. Handles the vast majority of tasks at 4-8x less cost than Pro. Supports thinking mode for reasoning tasks.

V4 Pro

The Heavy Lifter

Params1.6T total / 49B active
Context1M tokens
Max Output384K tokens
Price$0.435 / $0.87 per 1M
Cache Hit$0.0036 per 1M

Escalate when Flash fails: multi-step reasoning, competitive math, code-golf problems, long-horizon agent planning, heavy self-debugging coding sessions. Codeforces rating 3,206 - open-source SOTA in agentic coding.

03 The Cost Gap

DeepSeek is 10-34x cheaper than the US frontier.

At 10 million output tokens per month - a moderate enterprise workload - V4 Pro costs $8,700/month. GPT-5.5 costs $300,000/month for the same volume.

Model Input / 1M Output / 1M
DeepSeek V4 Flash $0.14 $0.28
DeepSeek V4 Pro $0.435 $0.87
Gemini 3.1 Pro $1.25 $5.00
GPT-5.4 $1.75 $8.00
GPT-5.5 $5.00 $15.00

All prices per 1M tokens, USD. DeepSeek pricing verified July 2026. Cache hit pricing (1/10th of standard input) not shown - available on both V4 models. May 2026: V4 Pro discount made permanent at 75% off standard rates.

04 Where to Run It

DeepSeek runs everywhere - API, cloud, on-prem, even your laptop.

API

Official

api.deepseek.com with OpenAI-compatible endpoints. Free chat at chat.deepseek.com. Both V4 models available. Concurrency: 2,500 (Flash), 500 (Pro). Routes through China-based infrastructure - consider latency for US/EU workloads.

Cloud

Managed

Available on Amazon Bedrock, Microsoft Azure AI Foundry, and Google Vertex AI. Also on data platforms: Snowflake Cortex AI and Databricks. OpenRouter provides multi-provider access with unified billing.

Self-Hosted

Open Weights

MIT licensed: download weights from Hugging Face, deploy on your own infrastructure. Day-zero support from vLLM, SGLang, and Hugging Face TGI. Distilled versions (1.5B to 70B) run on consumer hardware via Ollama or LM Studio.

05 Tradeoffs

The price is real. So are the tradeoffs.

Consider

What you get

Frontier-quality reasoning and coding at 10-34x less than GPT-5.5. Open weights under MIT license - no vendor lock-in, full self-hosting rights. 1M token context window without surcharges. Cache hit pricing at 1/10th of standard input. Active research pipeline shipping new models roughly every 6 months.

Watch for

What you trade

Official API routes through China - latency for US/EU users and data residency implications. Thinner SDK ecosystem vs OpenAI/Anthropic. Rate limits and reliability guarantees differ. No web browsing, no plugin ecosystem, no native multimodal. The cheapest model that can't meet your quality bar isn't cheap - benchmark before switching.

"The price gap is not a promotion. It is a structural advantage - open weights, MoE efficiency, and a research lab that trains frontier models for 1/20th the cost."

01

Default to V4 Flash. Escalate to V4 Pro only when Flash measurably fails on a specific task. Flash handles 80%+ of production workloads at 4-8x less cost.

02

Migrate off deepseek-chat and deepseek-reasoner aliases before July 24, 2026. The old names return errors after that date. One line of code: swap to deepseek-v4-flash.