Rapidly rising Chinese lab; known for efficiency and strong benchmarks at lower cost.
DeepSeek V4-Pro (1.6T parameters, 1M context window) rivals top closed-source models at a fraction of the cost, and is the first release optimized for Huawei Ascend chips; the Flash variant sacrifices some capability, and real-world agent performance is less proven than OpenAI or Anthropic models.— Claude Opus 4.6
DeepSeek's efficiency breakthroughs (sparse MoE, training on domestic chips) have been genuinely influential, with the V3/R1 papers widely cited and MIT Tech Review calling V4 a model that 'matters'; but the team is small and research breadth is narrow compared to Google or Meta.— Claude Opus 4.6
Offers an API platform with extremely competitive pricing ($0.145/M input on V4 Pro after a 75% promo) and open-weight model downloads; but no consumer chatbot with significant adoption, no enterprise platform, and limited developer tooling compared to OpenAI, Google, or Anthropic.— Claude Opus 4.7
Reported $20B funding round talks with Tencent and Alibaba validate market interest, V4's pricing is reshaping the API economics conversation, and the Huawei Ascend partnership lowers cost structure; revenue is still minimal and the business model is primarily open-source plus discounted API.— Claude Opus 4.7
Demonstrated remarkable efficiency by training frontier models on Huawei Ascend 950 'Supernode' clusters (reducing NVIDIA dependency) and V4 was first to be optimized for Ascend; total compute capacity is small, and the funding round is partly motivated by the need to scale infrastructure.— Claude Opus 4.7
DeepSeek releases model artifacts and technical information, which supports external testing, but its safety disclosures, governance details, and regulatory trust posture are limited; allegations around IP and national-security concerns further reduce trust.— GPT-5.5 Pro
A small but extraordinarily talented team has punched far above its weight, producing frontier-adjacent models with a fraction of competitors' resources; the funding round is explicitly motivated by retention concerns, and key-person risk is very high for a team this small.— Claude Opus 4.7
DeepSeek models are among the most downloaded on Hugging Face with massive open-source adoption, and the API is gaining traction; but there is no consumer product, limited enterprise presence, and distribution relies on third-party platforms.— Claude Opus 4.6
Efficiency innovations and training methodology represent a real technical moat that is hard to replicate; but the open-source strategy means competitors can build on DeepSeek's work, and the company lacks proprietary data sources.— Claude Opus 4.7
V4-Flash and V4-Pro launched April 24 with a 75% promo, the Huawei Ascend partnership shifted the China-US AI rivalry narrative (per CFR), and a $20B funding round is reportedly imminent; the asterisk is that distribution and product breadth still lag the model itself.— Claude Opus 4.7
Tracked for AI Power Rankings scoring. Covers model releases, benchmarks, pricing, funding, partnerships, infrastructure, and policy changes.
Published: 2026-04-24 | Logged: 2026-05-02T09:00Z | Area: Model Quality, Business & Market, Compute & Infra
DeepSeek launched V4 Flash and V4 Pro on April 24, 2026, both mixture-of-experts models with 1M token context windows. V4-Pro has 1.6T total / 49B active params; V4-Flash has 284B total / 13B active params. V4 Flash costs $0.14/M input tokens; V4 Pro costs $0.145/M input and $3.48/M output tokens, significantly undercutting competitors. A 75% promotional discount on V4 Pro ran through May 5. DeepSeek acknowledges V4 "trails state-of-the-art frontier models by approximately 3 to 6 months." DeepSeek also partnered with Huawei, using its Ascend 950 "Supernode" clusters for compute.
Scoring impact: Major boost to Model Quality (near-frontier performance). Strong Business & Market impact (aggressive pricing undercuts all competitors). Compute & Infra bolstered by Huawei partnership.
Sources:
Published: 2026-05-02 | Logged: 2026-05-03T18:00Z | Area: Model Quality, Business & Market, Safety & Alignment
The U.S. Center for AI Standards and Innovation (CAISI) at NIST published its evaluation of DeepSeek V4 Pro, calling it the most capable PRC AI model CAISI has tested to date. Across 16 benchmarks spanning cyber, software engineering, natural sciences, abstract reasoning, and mathematics (against 35 models), CAISI found V4's capabilities trail the frontier by approximately 8 months — performing similarly to GPT-5 (released ~8 months prior) rather than the more recent GPT-5.4 / Opus 4.6 that DeepSeek targeted. V4 was, however, more cost-efficient than GPT-5.4 mini on 5 of 7 benchmarks (53% cheaper to 41% more expensive depending on benchmark).
Scoring impact: Mildly negative on Model Quality vs DeepSeek's own claims (independent eval shows ~8-month lag). Reinforces Business & Market positioning on cost-efficiency. Safety & Alignment-relevant given CAISI's role in national-security-adjacent capability tracking.
Sources: