DeepSeek
Open-weight MoE models for code, math and agent loops. V4-Flash lists $0.14/M input tokens and ships MIT weights on Hugging Face.
Updated 2026-08-10
Yes, within limits. DeepSeek runs a free tier with paid plans above it.
Listed pricing: Free chat + API from $0.14/M input tokens (V4-Flash beta).
Overview
DeepSeek ships open-weight models aimed at code, math and multi-step reasoning, and sells API access to them at rates that sit an order of magnitude under the frontier tiers. The flagship is V4-Pro, a 1M-context model whose 75% price cut became permanent on May 23, 2026. Alongside it sits DeepSeek-V4-Flash, opened to public beta on July 31, 2026 under the dated model ID deepseek-v4-flash-0731: 284B total parameters with 13B active, a 1M-token context window, and MIT-licensed weights on Hugging Face.
The V4-Flash release is where the 2026 story actually is, because it added a native Responses API surface next to the Chat Completions endpoint DeepSeek has served for years. On that surface the server holds the turn history and the client passes a response id forward, so a ten-step agent run stops resending its step-one context ten times, and the model's reasoning items can survive a tool call. DeepSeek also states the model is "fully adapted for Codex," which means Codex CLI and similar harnesses can register it as a custom provider without a chat shim. Artificial Analysis listed the beta at $0.14 per million input tokens, $0.28 output and $0.003 on cache hits as of August 2, 2026, with a blended $0.06 on a 7:2:1 mix of cache hits to fresh input to output, and scored the reasoning variant at 50 on its Intelligence Index against a 25 median, third of 101 large open-weights models.
The honest take: the cheap number is conditional on your code. That 98% cache discount only applies to a stable prefix, so a harness that reshuffles tool definitions or injects a timestamp near the top of the system prompt pays full freight, with an output leg Artificial Analysis measured at roughly twice the median length. The envelope has other hard edges: text in, text out with no multimodal input, no published throughput figure, a beta endpoint with no SLA, and China-hosted inference that now holds your conversation state server-side if you use the Responses path. DeepSeek was the disruptive outlier in 2025; in 2026 it shares that shelf with Qwen, GLM, MiniMax and Kimi, and it wins on price discipline and license terms rather than on being the only cheap option.
Is DeepSeek free?
Yes, within limits. DeepSeek runs a free tier with paid plans above it.
What the free tier covers: Free web and mobile chat with daily usage limits; API is pay-as-you-go, check the platform console for any current credit offer.
Listed pricing: Free chat + API from $0.14/M input tokens (V4-Flash beta).
Pricing verified . That is the date the plans were last re-checked against the vendor's own pages, not today's date. Confirm on the official site before you pay.
What the free tier leaves out
Read straight off the plan list below. Vendors move features between tiers, so check the current split before you pay.
- API — DeepSeek-V4-Flash (beta) $0.14/M input, $0.28/M output Cache hits $0.003/M. 1M context. Rates as listed by Artificial Analysis, August 2, 2026
- API — DeepSeek-V4-Pro 75% below original rates, permanent since May 23, 2026 Flagship 1M-context model. Per-token figures vary by cache, context and batch mode — check the current pricing page
DeepSeek pricing
| Plan | Price | What's included |
|---|---|---|
| Free Chat | Free | Web and mobile chat with daily usage limits |
| API — DeepSeek-V4-Flash (beta) | $0.14/M input, $0.28/M output | Cache hits $0.003/M. 1M context. Rates as listed by Artificial Analysis, August 2, 2026 |
| API — DeepSeek-V4-Pro | 75% below original rates, permanent since May 23, 2026 | Flagship 1M-context model. Per-token figures vary by cache, context and batch mode — check the current pricing page |
Web and mobile chat with daily usage limits
Cache hits $0.003/M. 1M context. Rates as listed by Artificial Analysis, August 2, 2026
Flagship 1M-context model. Per-token figures vary by cache, context and batch mode — check the current pricing page
Pricing verified . That is the date the plans were last re-checked against the vendor's own pages, not today's date. Confirm on the official site before you pay.
Is DeepSeek worth it?
Worth it for Developers who own their agent harness and can hold a stable prompt prefix, since the cache rate is what makes DeepSeek cheap, and whose work is text rather than screenshots or audio.
You can test that on the free tier before paying anything. The recorded trade-offs are listed below, and any one of them can settle the question on its own.
The 8.7/10 AI Score is an editorial read of published capability, price and shipping pace. Nobody here has hands-on hours with DeepSeek. How we verify.
Worth it if
The strengths recorded against this entry.
- V4-Flash lists $0.14/M input and $0.28/M output, roughly 36x and 107x under GPT-5.6 Sol's $5 and $30
- Cache hits at $0.003/M make a repeated agent prefix nearly free, blending to about $0.06/M on an agent-shaped traffic mix
- MIT-licensed 284B/13B-active weights on Hugging Face, so the same model can be self-hosted or fine-tuned
- 1M-token context on both V4-Pro and V4-Flash handles whole codebases without chunking
- Native Responses API plus stated Codex adaptation lets agent harnesses point at it without a chat shim
Not worth it if
Any one of these blocks your use case.
- Text in, text out — no image, audio or video input, so screenshot-driven debugging is off the table
- V4-Flash is a public beta: no SLA, movable rate limits, and no published throughput figure (Artificial Analysis lists Speed as N/A)
- Verbose output — 210M tokens across nine evaluations against a 100M median, which eats into the cheap-token advantage
- China-hosted inference, and the Responses path parks conversation state on DeepSeek's servers, a compliance question before an engineering one
What sets DeepSeek apart
- Cache-hit input at $0.003 per million, a 98% discount on its own list rate and ranked first in its size class by Artificial Analysis
- MIT-licensed weights for a 284B-total / 13B-active MoE published on Hugging Face, so the identical model can be self-hosted
- Native Responses API from a non-OpenAI provider, plus stated Codex CLI adaptation
- 1M-token context on both V4-Pro and V4-Flash at open-weights pricing
Key features
Agent-Native API
The V4-Flash beta serves a native Responses API next to Chat Completions. The server keeps turn history behind a response id, so agent loops send only new items and reasoning can persist across a tool boundary. DeepSeek states the model is fully adapted for Codex.
Cache-Priced Context
1M-token context with cache hits listed at $0.003 per million against $0.14 for fresh input. Artificial Analysis quotes a blended $0.06 per million on a 7:2:1 agent-shaped mix, which only holds if your prompt prefix stays stable between turns.
MIT Open Weights
V4-Flash is a 284B-parameter mixture-of-experts with 13B active per token, released under MIT on Hugging Face. Self-hosting is a hardware decision before a licensing one: Unsloth's full-precision Q8 build runs 162GB.
Reasoning + Code
Artificial Analysis scores the reasoning variant at maximum effort at 50 on its Intelligence Index, third of 101 large open-weights models. Three of the nine evaluations in that composite are agentic: GDPval-AA v2, tau-3 Banking and Terminal-Bench v2.1.
How it compares
| Tool | Best for | Pricing | Score |
|---|---|---|---|
| DeepSeek | Developers who own their agent harness and can hold a stable prompt prefix, since the cache rate is what makes DeepSeek cheap, and whose work is text rather than screenshots or audio. | Free chat + API from $0.14/M input tokens (V4-Flash beta) | 8.7/10 |
| Claude Code vs Claude Code → | Developers who want an agent that works inside an existing repo and toolchain rather than in a hosted editor, and who already pay for a Claude plan. | Included with Claude Free, Pro, Max, Team and Enterprise plans | 9.3/10 |
| Cursor vs Cursor → | Professional developers who want an IDE-native agent that can also run in the cloud and be steered from a phone, and engineering teams that need admin control over which model handles which request. | Free Hobby + Individual $20/mo + Teams $40/user/mo + Enterprise custom | 9.2/10 |
| v0 by Vercel vs v0 by Vercel → | Frontend developers on React and Next.js who want production-ready UI generated from a text prompt or a screenshot instead of hand-built boilerplate. | Free tier + Plus $30/user/mo + Business $100/user/mo + Enterprise custom | 9/10 |
Learn to use DeepSeek
Step-by-step, each one stamped with the date it was last checked against the docs.
Compare head-to-head
Related reading
DeepSeek V4-Flash vs GPT-5.6 Sol: $0.14 Against $5
DeepSeek's V4-Flash beta lists $0.14 and $0.28 per million tokens against GPT-5.6 Sol's $5 and $30. The cache line and the agent wire format decide the bill.
DeepSeek V4 Flash 0731's Score of 50, Audited
Artificial Analysis scores DeepSeek V4 Flash 0731 at 50 on its Intelligence Index. What the number measures, what it leaves out, and what it costs per task.
DeepSeek-V4-Flash API: Responses, Codex, Agent Setup
DeepSeek opened the V4-Flash public beta on July 31, 2026 with a native Responses API. Model ID, server-side state, Codex setup, and the agent claim.
Ready to try DeepSeek?
Head to the official site to start with DeepSeek — pricing and plans are listed above.
Visit DeepSeek