Summary
- • DeepSeek offers 75% discount on V4-Pro API until 5 May 2026, cutting input tokens to ~$0.036 per million.
- • At full price, V4-Pro already undercuts GPT-5.5, Claude Opus 4.7, and Gemini 3.1 Pro; cache-hit pricing cut to 1/10th across entire API suite.
- • V4-Pro is the largest open-weight model at 1.6T parameters, trained on Huawei Ascend 950 chips — not Nvidia GPUs.
- • Move extends the pricing pressure DeepSeek first triggered in January 2025 with its R1 model.
Details
DeepSeek's playbook: open-source removes model access barrier; aggressive pricing removes cost barrier
The combination systematically lowers every friction point for switching from OpenAI, Anthropic, or Google APIs. V4-Pro integrates natively with Claude Code, OpenClaw, and OpenCode — the dominant agentic coding frameworks — so developers can substitute DeepSeek's API without changing their toolchain. A 1M-token context window adds enterprise viability for large-codebase and long-document use cases requiring a single API call.
1.6T-parameter MoE trained on Huawei Ascend 950 and Cambricon hardware — not Nvidia
Training on domestic Chinese chips rather than Nvidia GPUs demonstrates a viable path to frontier-class AI development outside the Nvidia supply chain. Counterpoint Research analyst Wei Sun noted this allows AI systems to be built and deployed without relying solely on Nvidia, with implications for both domestic Chinese adoption and broader AI hardware diversification globally.
White House accused Chinese firms of industrial-scale distillation of American AI models
White House Director of Science and Technology Policy Michael Kratsios accused foreign entities primarily based in China of conducting industrial-scale capability extraction from American AI models. DeepSeek's aggressive pricing announcement arrives amid heightened US-China tensions over AI development practices and technology transfer.
Continues pricing disruption DeepSeek launched in January 2025 with R1 model
R1 claimed frontier-level reasoning performance at a fraction of OpenAI's costs, sparking an industry-wide pricing reassessment. V4-Pro continues that trajectory as the largest open-weight model yet, creating sustained downward pressure on US AI labs' API pricing at a moment when those labs are simultaneously investing heavily in compute infrastructure.
iiMedia's Zhang Yi: V4's architecture is a 'genuine inflection point' for long-context AI processing
Predicts ultra-long context support will move from research settings into mainstream commercial applications. The combination of 1M-token context, cache pricing cuts, and open-weight availability positions V4-Pro to accelerate enterprise adoption for workloads previously requiring expensive multi-call workarounds.
Strategy = pricing/positioning logic, Tech Info = architecture and hardware, Context = geopolitical background, Market Impact = competitive dynamics, Insight = analyst perspectives
What This Means
For developers and startups whose primary constraint is API cost, DeepSeek has made the switching calculus from US AI providers straightforward — lower price, comparable or superior specs, and no toolchain disruption. US AI labs face sustained downward pressure on API pricing at a moment when they are simultaneously investing heavily in compute infrastructure. The broader implication is that DeepSeek has demonstrated frontier-class model development, open-weight release, and aggressive commercial pricing are all achievable on non-Nvidia hardware, challenging assumptions that have underpinned Western AI industry strategy.
Sentiment
Broadly excited about affordability and price competition, with analysts noting pressure on US AI labs
“Pricing is becoming everything in AI. The winning model will have Opus like performance and ideally be priced 10x less! DEEPSEEK IS COMING”
“deepseek is starting a price war on the ai market... ~17x cheaper than gpt-5.5 and ~14x cheaper than opus 4.7... ai is starting to commodify. the price war has begun”
“Lower cost than frontier models... but high token usage keeps costs above most open weights peers... DeepSeek V4 Pro hallucinates 94% of the time when it doesn't know an answer”
Detailed benchmark leader but highlights practical limitations
“Excited to see the impact of the innovations... Engram Conditional Memory... mHC... DeepSeek Sparse Attention... enabling massive models more economically feasible”
“DeepSeek is about to release V4... run natively on Huawei silicon... China has closed the loop... competitive dynamic shifting... pressure on NVIDIA”
Emphasizes geopolitical and hardware independence implications
Split
~80/20 excited about lower costs for devs vs analytical notes on US labs' pricing pressure and model gaps.