Open-weight LLM matching GPT-4 quality at a fraction of the API cost. Strong code and math performance, MIT-licensed weights for self-hosting.
AI Assistant
DeepSeek V3 is the open-weight, mixture-of-experts (MoE) language model family from Hangzhou-based DeepSeek AI, the lab that spent 2025 forcing every proprietary frontier vendor to explain its per-token pricing. The line spans the original V3 (December 2024), V3-0324 (March 2025, 685B parameters), V3.1 (August 2025), V3.1-Terminus, and V3.2 (released December 1, 2025) — the last of which introduced DeepSeek Sparse Attention (DSA) and scored 73.1% on SWE-Bench Verified, roughly the same neighborhood as Claude Sonnet 4.5 at a small fraction of the sticker price. All V3 checkpoints share the same core: 671B total parameters with 37B active per token, 128K–164K context, and full model weights available on Hugging Face under a permissive license.
The honest pitch: if you are a solo builder, a small team shipping an agent, or a mid-market engineering group running a high-volume inference workload, DeepSeek V3 is the cheapest way to get GPT-5.4-mini-class output quality without giving up the ability to self-host. It is not the smartest model on the market — DeepSeek's own V4-Pro replaced it as the flagship in April 2026, and Claude Sonnet 5 clears it on complex agentic coding — but it is the model most likely to survive a finance review when your app crosses 10M tokens per day. If your budget is real and your workload is standard (chat, summarization, code completion, retrieval-augmented generation, batch classification), V3.2 is close to a default answer.
The competitive frame in August 2026 is crowded and it matters. Qwen 3.7 Flash undercuts DeepSeek on raw input price ($0.03/M vs $0.21/M for V3.2) but loses on reasoning benchmarks. Llama 4 Scout on Together.ai runs $0.08/$0.18 with better tool-calling but a shorter context. Mistral Small 4 offers EU data residency at $0.15/$0.60 for regulated workloads DeepSeek cannot legally serve. And Claude Sonnet 5 at $2/$10 (introductory through Aug 31, 2026) is the model you buy when quality wins over cost. DeepSeek V3's specific differentiator is the combination of frontier-grade reasoning benchmarks, open weights you can actually run yourself, and a hosted API that costs less than a coffee per million tokens.
671B-parameter Mixture-of-Experts with 37B active. V3 uses 256 routed experts with 8 activated per token plus one shared expert, which is how it delivers 671B-parameter quality at 37B-parameter inference cost. In practice this means the hosted API responds at latencies competitive with dense 30–40B models like Llama 3.3 70B or Mistral Medium, but with the depth of a frontier model on hard reasoning tasks. The architecture is what makes the pricing arithmetic work — you are not subsidizing 671B of compute per token.
DeepSeek Sparse Attention (V3.2 only). The December 2025 release introduced a fine-grained sparse attention mechanism that reduces long-context inference cost without measurably degrading quality up to 128K tokens. If you are running long-document RAG, agentic workflows with big scratchpads, or codebase-wide analysis, V3.2 is meaningfully cheaper to run at 100K+ context than V3-0324 or any dense competitor. This is the single largest reason to prefer V3.2 over the older V3 checkpoints.
Open weights under a permissive license. Full model weights are downloadable from Hugging Face and GitHub. You can fine-tune, quantize, distill, or self-host — none of which is possible with GPT-5, Claude, or Gemini. Running the full 671B model requires roughly 8x H100s or equivalent (~$25K/month on cloud), so self-hosting only makes economic sense at high volume. But you own the model, and you can walk away from the hosted API the day China export controls or your privacy team tell you to.
Controllable reasoning mode. V3.1 and later support a thinking mode that can be toggled per request. Reasoning-on gives you chain-of-thought output with visible reasoning tokens (billed at output rates), reasoning-off returns a direct answer at lower cost. This is the same pattern OpenAI charges a separate SKU for (o-series) — DeepSeek folds both into one model and one price. Practical impact: you can route hard problems to reasoning mode and cheap ones to direct mode without switching endpoints.
73.1% SWE-Bench Verified on V3.2. V3.2 lands within a few points of Claude Sonnet 4.5 on the industry-standard software engineering benchmark, and V3.2-Speciale (the high-compute reasoning variant) hits 96% on AIME 2025. Gold-medal placements on IMO 2025 and IOI 2025 are officially documented. This is not marketing math — these benchmarks are the same ones every proprietary vendor cites. On code and math, DeepSeek is not a discount option; it is a peer.
Cache pricing that actually saves money. The hosted API charges roughly 10% of input rate on cache hits — meaningful for agentic workloads that repeatedly send the same system prompt or tool schema. On V3.2 that works out to cache-hit input at roughly $0.02/M vs $0.21/M cold. If your app has any prompt structure that repeats (and most do), your effective input bill drops 60–80% once caching warms up. Anthropic charges a similar structure; OpenAI's cache is auto-managed and less predictable.
Multi-provider hosting. Beyond DeepSeek's own API, V3 checkpoints are available on OpenRouter, Together.ai, Fireworks, DeepInfra, Novita, and Chutes. Prices vary — Together and Fireworks are typically 10–30% above DeepSeek's direct API but ship with US data residency, higher rate limits, and enterprise SLAs. This is the workaround for teams that need V3's quality but cannot send data to a China-jurisdiction endpoint.
128K–164K context on standard tiers. No premium upcharge for long context. Compare to Claude Sonnet 5 (200K standard, 1M behind a Tier 4 gate) and GPT-5.5 (1M context but at $5/$30 per M). For long-context work at scale, V3.2 is by a wide margin the cheapest option that isn't Gemini 2.5 Flash.
DeepSeek publishes flat per-token rates with no minimums, no seat fees, and no enterprise gate. Cache-hit pricing is automatic. Off-peak billing (roughly 50% off, UTC 16:30–00:30) applies to all models on the direct API.
| Tier | Input (per 1M) | Output (per 1M) | Cache Hit | Context |
|---|---|---|---|---|
| DeepSeek V3 (base) | $0.27 | $1.10 | $0.07 | 128K |
| DeepSeek V3-0324 | $0.24 | $0.90 | $0.024 | 164K |
| DeepSeek V3.1 | $0.25 | $0.95 | $0.025 | 128K |
| DeepSeek V3.2 (current) | ~$0.21 | ~$0.32 | ~$0.02 | 164K |
| DeepSeek V3.2-Speciale (reasoning) | ~$0.28 | ~$1.10 | ~$0.03 | 164K |
V3.2 rates reflect the standing 25% discount promoted through Q3 2026; DeepSeek may return them to nominal ($0.28/$0.42) after the promo window closes. [verify pricing at deepseek.com/pricing before committing production spend.]
For a solo creator or a small team building an agent, the arithmetic is stark. A workload that consumes 50M input and 20M output tokens per month costs roughly $17 on V3.2, roughly $250 on Claude Sonnet 5 introductory, roughly $600 on Claude Sonnet 4.5, and roughly $850 on GPT-5.5. Even against DeepSeek's own V4-Flash ($0.14/$0.28), V3.2 is only marginally more expensive on input and cheaper on output — and V3.2 wins outright on SWE-Bench. The one honest comparison where V3 loses on price is Qwen 3.7 Flash at $0.03/$0.13, but Qwen trails on code generation and long-context reasoning by a wide enough margin that most engineering teams pay the DeepSeek premium.
Pros
Cons
Qwen 3.7 Flash / Qwen 3.8 Max ($0.03–$2 input / $0.13–$6 output). Alibaba's open-weight family. Flash is the price floor for open weights and wins on Chinese and multilingual work; Max competes on reasoning. Pick Qwen if your workload is dominated by translation, Asian languages, or extreme cost sensitivity where a 5–10% quality gap on English coding is acceptable. Pick DeepSeek V3 for English/code work where SWE-Bench and AIME scores actually matter.
Llama 4 Scout / Maverick (~$0.08–$0.60 per M via Together/Fireworks). Meta's open-weight line. Strongest tool-calling in the open-weight category and best-supported by fine-tuning tooling (Unsloth, Axolotl, Together fine-tunes). Pick Llama 4 if you plan to fine-tune heavily or need the widest ecosystem of quantized and specialized variants. Pick V3 for out-of-the-box reasoning and lower total cost at scale without fine-tuning.
Claude Sonnet 4.5 / Sonnet 5 ($2–$3 input / $10–$15 output). Anthropic's mid-tier. Better agentic behavior, better instruction following, cleaner refusals, mature tool use, US jurisdiction. Pick Sonnet when the workload is user-facing, when quality per token matters more than cost per token, or when you need Claude Code / MCP ecosystem integration. Pick DeepSeek V3 when a 5–8% quality delta will save you $500+/month.
GPT-5.4-mini / GPT-5.4-nano ($0.20–$0.75 input / $1.25–$4.50 output). OpenAI's cost tier. Better vision, native voice, official SDK ecosystem, ChatGPT brand recognition for consumer apps. Pick GPT-5.4 for multimodal or when the client wants "we use OpenAI." Pick DeepSeek V3 for pure text/code where cost dominates.
DeepSeek V4-Flash / V4-Pro ($0.14–$0.435 input / $0.28–$0.87 output). Same vendor's successor. V4 is smarter on the frontier tasks V3 struggles with (long-horizon agents, novel reasoning), but V3.2 is still cheaper on output and roughly matches on standard workloads. Pick V4 if you are starting fresh and want the current flagship; keep V3.2 if it is already working in production and the cost delta matters.
Buy DeepSeek V3.2 if you are a solo builder, a small engineering team, or a mid-market shop running a high-volume text or code workload where cost dominates the decision and your data can legally leave China (or you can route to Together / Fireworks / self-host). Buy V3.2-Speciale if you specifically need frontier reasoning at a fraction of Opus or GPT-5.5 pricing. Skip V3 for user-facing production apps that need mature agentic tool use (buy Claude Sonnet 5), for multimodal work (buy GPT-5.4 or Gemini 2.5), for regulated data (buy Mistral or self-host), and for greenfield builds where the small V4 premium is worth having the current flagship. Pick DeepSeek V3.2 when you have measured your token spend and can no longer justify the proprietary markup; skip it when brand, jurisdiction, or multimodality is on the line.
Q: Is DeepSeek V3 free? A: DeepSeek V3 model weights are free to download and self-host under a permissive license from Hugging Face and GitHub. The hosted API is not free but is priced at roughly $0.21/M input and $0.32/M output for V3.2 as of August 2026 — cheap enough that light experimentation costs cents. There is no perpetual free API tier; new keys get modest starter credits.
Q: How does DeepSeek V3 compare to Claude Sonnet 4.5? A: DeepSeek V3.2 scores 73.1% on SWE-Bench Verified vs Sonnet 4.5's 77.2%, at roughly one-fifteenth the input cost and one-fortieth the output cost. Sonnet wins on agentic tool use, refusal quality, and Claude Code ecosystem. V3 wins on price, open weights, and long-context economics. For batch workloads pick V3; for user-facing agents pick Sonnet.
Q: Can I self-host DeepSeek V3? A: Yes — full weights are on Hugging Face. The full 671B MoE needs roughly 8x H100 or equivalent (~$25K/month cloud) at FP8. Quantized Q4/Q5 versions run on 2–4 H100s or a well-specced H200 box with measurable but usually acceptable quality loss. Below that, use distilled variants or the hosted API.
Q: Is the DeepSeek API safe for enterprise data? A: The direct api.deepseek.com endpoint is hosted in China and subject to Chinese jurisdiction. For US, EU, or regulated workloads route through Together.ai, Fireworks, DeepInfra, or self-host — same weights, US or EU data residency, roughly 10–30% price premium.
Q: Should I use DeepSeek V3 or V4? A: Use V4-Pro if you need the current flagship, want the latest agent capabilities, and are building greenfield. Use V3.2 if you have an existing deployment that works, are cost-sensitive on output tokens, or need V3.2-Speciale's reasoning benchmarks. V4-Flash and V3.2 overlap in price and capability — V3.2 is stronger on reasoning, V4-Flash is stronger on agentic tool use.
Q: What context window does DeepSeek V3 support? A: 128K tokens on base V3 and V3.1, 164K on V3-0324 and V3.2, no premium gate at any tier. Sparse attention on V3.2 makes long-context inference roughly 2–3x cheaper than V3-0324 at the same context length.
Q: Does DeepSeek V3 support tool use and structured output? A: Yes — native function calling, JSON mode, and structured outputs are supported on all V3 checkpoints from V3.1 forward. Tool-use quality trails Claude Sonnet and GPT-5.4 but matches Llama 4 and beats Qwen 3 in most agent benchmarks.
Q: Is DeepSeek V3 good for coding? A: Yes — 73.1% SWE-Bench Verified on V3.2 puts it in the same tier as Claude Sonnet 4.5 and ahead of GPT-5.4-mini. It is a competent daily driver in Cursor, Cline, Aider, and Continue.dev, and it is the default cost-optimized backend for many agentic coding setups in 2026.