DeepSeek launched V4 Flash and V4 Pro open-weight models with 1M token context, claiming near-parity with GPT-5 and Gemini 3.x at drastically lower prices.
DeepSeek released two preview versions of DeepSeek V4: a Flash variant (284B total / 13B active parameters) and a Pro variant (1.6T total / 49B active parameters), making the Pro the largest open-weight model available. Both use mixture-of-experts architecture with 1 million token context windows. Pricing is set at $0.145/M input tokens and $3.48/M output tokens — undercutting OpenAI's GPT-5.4, Gemini 3.1 Pro, and Claude Opus 4.7. DeepSeek claims near-frontier performance on reasoning and coding benchmarks, though it trails on knowledge-intensive tasks. The launch coincides with U.S. accusations of Chinese AI IP theft at industrial scale.
DeepSeek V4 Pro's 49B active parameters (MoE) with a 1M token context window means you can run massive codebase-aware completions at a cost that makes GPT-5.x look expensive. The Flash variant at 13B active params is a serious candidate to replace mid-tier proprietary models in latency-sensitive pipelines. Benchmark parity on coding tasks with GPT-5.4 is the headline; the knowledge-task gap is real but narrow for most production use cases.
Swap your current GPT-4-class API calls to DeepSeek V4 Flash on a single high-volume endpoint this week and benchmark cost-per-request against your current spend — if throughput holds, you're looking at an immediate 70–85% cost reduction on inference.
Get your DeepSeek API key from platform.deepseek.com and export it: export DEEPSEEK_KEY=your_key_here
Tags
Sources
Related