DeepSeek releases V4-Pro (1.6T params, 49B active) and V4-Flash (284B, 13B active) with MIT license, undercutting all frontier model pricing.
Key Takeaways
V4-Flash costs $0.14/M input, $0.28/M output – cheaper than GPT-5.4 Nano; V4-Pro at $1.74/$3.48 undercuts Gemini 3.1 Pro and Claude Sonnet 4.6.
Both models support 1M token context; V4-Pro achieves only 27% single-token FLOPs and 10% KV cache size vs DeepSeek V3.2 at 1M context.
DeepSeek self-reports V4-Pro-Max trailing GPT-5.4 and Gemini-3.1-Pro by roughly 3-6 months on reasoning benchmarks.
V4-Pro at 865GB is the largest open-weights model released, exceeding Kimi K2.6 (1.1T) and GLM-5.1 (754B); Flash at 160GB may run quantized on 128GB M5 MacBook Pro.
The current API price includes a 75% discount valid through 2026-05-31, meaning post-discount pricing will be roughly 4x higher.
Hacker News Comment Review
The 75% API discount is active until May 31, 2026 and was not factored into the article’s pricing table – real post-promo costs change the competitive picture significantly.
Practical coding evaluations rate V4-Pro as Sonnet-level for implementation tasks but below Opus 4.7 max thinking for open-ended planning; a max thinking mode does exist in the API.
Data privacy risk of routing production codebases through DeepSeek’s API (direct or via OpenRouter) drew little community pushback despite analogous concerns raised about GitHub Copilot training data.
Notable Comments
@holysantamaria: Flags that the pricing page lists a 75% discount on V4-Pro extended until 2026-05-31 – not reflected in the article’s comparison table.
@zozbot234: Notes V4-Pro has a max thinking mode in the API, relevant for direct capability comparisons with Opus 4.7.
