DeepSeek V4 makes 1M context standard across all models
DeepSeek today released a preview of its new V4 series models, open-sourcing them alongside. The biggest change: 1 million tokens of context is now standard across all official services, regardless of version or price tier.
V4 comes in two flavors: V4-Pro, the flagship, and V4-Flash, the lightweight edition. According to DeepSeek's own benchmarks, V4-Pro's reasoning ability matches top-tier closed-source models, and its world knowledge trails only Gemini-Pro-3.1.
What's interesting is that DeepSeek proactively compared its model to Anthropic's: internal employees using V4-Pro for agentic coding (letting AI autonomously complete programming tasks) reported a better experience than Claude Sonnet 4.5, with delivery quality close to Opus 4.6's non-thinking mode — but still lagging behind Opus 4.6 with deep thinking enabled. This kind of "openly acknowledging the gap" is rare in Chinese companies' launch announcements, and to some extent it shows that Opus 4.6's thinking mode has become the industry's invisible ceiling.
V4-Flash is positioned as the budget-friendly option: reasoning ability is close to Pro, but it has less world knowledge and falls short on complex agent tasks. It's sufficient for most everyday scenarios and comes with a more attractive API price.
On the technical side, V4 introduces a new attention mechanism that compresses at the token level, paired with a self-developed DSA sparse attention, dramatically reducing the compute and memory requirements for 1M context. In simple terms: 1M context used to be "possible but expensive" — now it's "standard and no extra charge." For developers, this means you can throw an entire codebase or full documentation set into the model at once without worrying about chunking.
Another practical detail: V4 is specifically optimized for mainstream agent tools like Claude Code and OpenClaw. The API supports both OpenAI and Anthropic interface formats — switching only requires changing the model parameter. The old deepseek-chat and deepseek-reasoner endpoint names will work for another three months, then stop on July 24. Developers should migrate ahead of time.