DeepSeek V4 Makes Million-Token Context the Default

DeepSeek also claims major efficiency gains for long-context inference. The model card says V4-Pro at 1M context uses 27% of single-token inference FLOPs and 10% of KV cache compared with DeepSeek-V3.2. Treat that as a company claim until independent testers reproduce it, but the claim explains the product strategy: DeepSeek wants memory to look like an operational feature, not a demo stunt.

The caveat is the story

The strongest reason to take the launch seriously is also the reason not to overread it. DeepSeek says V4-Pro performs better than current open models on several reported evaluations, but those are still lab-reported results. Builders should wait for outside runs before treating the table as settled.

DeepSeek’s own agentic-coding comparison is more revealing. In its announcement, the company says internal feedback puts V4-Pro ahead of Claude Sonnet 4.5 and near Claude Opus 4.6 in non-thinking mode, while still behind Opus 4.6 in thinking mode. That caveat is the sentence the launch copy cannot quite hide.

A million tokens of context can let a model carry more of a repo and its surrounding documentation. It does not magically make the model better at planning or multi-step coding judgment. DeepSeek’s thinking-mode docs point to the same split by exposing reasoning_effort controls for V4-Pro, which makes the product feel less like a single model and more like a menu of reasoning tradeoffs across latency and cost.

There is one near-term migration detail worth watching. DeepSeek says older API aliases, including deepseek-chat and deepseek-reasoner, will stop being used on July 24, 2026. That gives existing customers a clock, not just a shiny model name.

For builders, V4’s real test is not whether DeepSeek can publish a stronger leaderboard row this week. It is whether a 1M-token window becomes reliable and cheap enough that teams stop architecting around context scarcity. If that happens, the frontier shifts from who can answer the prompt to who can carry the whole workspace without making the bill or latency unusable.

Apr 21, 2026