DeepSeek-V4: a million-token context that agents can actually use

The meaningful detail here isn't the number itself but what it signals about DeepSeek's positioning: a million-token context window is only useful if the inference cost is low enough for agents to actually call it repeatedly, and DeepSeek has historically competed on cost efficiency where Western frontier labs have not.

This lands the same day TechCrunch reported that Meta is acquiring millions of Amazon's custom CPUs specifically for agent workloads, suggesting the industry is converging on a shared assumption: agents will need to process far more context per call than current deployments support. The chip procurement story and this release are essentially two sides of the same bet. Meanwhile, the April 17 tokenmaxxing coverage raised a pointed concern that developers chasing raw throughput are generating maintenance debt rather than productivity gains. A million-token window doesn't resolve that tension; it arguably intensifies it by making it cheaper to be sloppy about what gets stuffed into context.

Watch whether Cursor, currently raising at a $50B valuation on enterprise momentum, integrates DeepSeek-V4's context window into its core product within the next two quarters. If it does, that's a signal that long-context inference is becoming a baseline expectation for developer tooling, not a premium feature.