DeepSeek
Frontier-grade open-weight models at a tenth of US prices — but it's an engine, not a platform, and the hosted service's China data residency is a hard no for sensitive work.
Element scores
Strengths
Frontier-quality models under MIT license at roughly a tenth of US API prices: V4-Pro (1.6T MoE, 49B active) and V4-Flash both ship a native 1M-token context with sparse attention, thinking modes and tool calls, plus disk-backed context caching that drops cache-hit input to $0.0028/1M. The R1-Distill ladder scales down to a single GPU, and OpenAI- plus Anthropic-format endpoints mean it drops into existing harnesses — including Claude Code — with a base-URL change.
Honest dings
The hosted app and API store prompts in China under a legal framework you cannot contract around — South Korea's regulator documented undisclosed transfers, and US federal agencies, multiple US states, Australia and Taiwan have banned it from government devices. The platform around the models is bare: no embeddings, vector store, fine-tuning service, evals, moderation API or MCP. Models are text-only, and model names shift under you (deepseek-chat and deepseek-reasoner retire 24/07/2026).
Sources
Every audit lists the research it rests on — transparency and traceability are the product. Tools evolve: each audit is a snapshot of its audit date, and re-audits supersede older versions (kept below for reference).
- api-docs.deepseek.com/quick_start/pricing — Official V4-Flash/V4-Pro pricing, 1M context, 384K max output, thinking modes, tool calls, deepseek-chat/reasoner deprecation 24/07/2026 (accessed 2026-07-10)
- api-docs.deepseek.com/news/news260424 — Official V4 Preview release note (24/04/2026): open-sourced, agent-capability optimizations, long-context efficiency (accessed 2026-07-10)
- teamai.com/blog/large-language-models-llms/understa… — Independent June 2026 lineup guide: V4-Pro 1.6T/49B active, V4-Flash 284B/13B, 32T training tokens, MIT licensing from V3 forward, R2 unreleased (accessed 2026-07-10)
- linkedin.com/pulse/deepseek-v32-2-most-intelligent-… — Artificial Analysis benchmark read: V3.2 (Reasoning) index 66 — level with Kimi K2 Thinking, ahead of Grok 4 and Sonnet 4.5 Thinking; Speciale lacks tool calling (accessed 2026-07-10)
- magazine.sebastianraschka.com/p/technical-deepseek — Independent technical analysis of V3 → V3.2 architecture and DeepSeek Sparse Attention (accessed 2026-07-10)
- witness.ai/blog/deepseek-security-concerns — China data residency per DeepSeek's own privacy policy, National Intelligence Law exposure, South Korea PIPC finding of undisclosed transfers (accessed 2026-07-10)
- conference-board.org/research/ceo-center-newsletter… — US federal agency and state government device bans; pending federal legislation (status July 2026) (accessed 2026-07-10)
- longbridge.com/en/news/271582811 — SCMP report (06/01/2026): interleaved-thinking deep research added to the app, app still lacks multimodal support, MAU trajectory (accessed 2026-07-10)
- aicpb.com/product/DeepSeek-AI-%E6%99%BA%E8%83%BD%E5… — Third-party MAU tracker: 138.98M app MAU April 2026 (vendor discloses no official figures) (accessed 2026-07-10)
- fireworks.ai/blog/deepseek-models — Independent production-deployment guide: deepseek-chat/reasoner remapping to V3.2 in late 2025, production caveats on the lineup (accessed 2026-07-10)