DeepSeek V4 — The model I already use and love

$0.14 per million tokens. 1 million context tokens. 284B total parameters, 13B active. Open weights with MIT license. DeepSeek V4 Flash isn't the most powerful model on the market, but it has the best price-performance ratio available today. While some are putting 220,000 GPUs to work, others are trying to unionize entire labs. And in between, you and I are trying to figure out what to use today to stay ahead. I'm going to sort all this out from my point of view. No hype.

DeepSeek V4 — The model I already use and love

Let's start with what hits closest to home. DeepSeek V4 launched on April 24 as open weights with an MIT license. And it's massive.

V4-Pro has 1.6T total parameters (49B active) with MoE architecture. V4-Flash, which I'm already using daily, has 284B total, 13B active, and costs $0.14/M input tokens. To give you an idea, that's a fraction of what equivalent models cost.

1M context tokens by default. On both models.

And while DeepSeek democratizes access to the best model on the market, other news shows the sector isn't just advancing — it's reconfiguring.

Safety, unions, and $160 million for startups

The UK AISI published its frontier model evaluation this week. The conclusion: the most advanced models already operate at expert-level cyberattack capability. This isn't futurism — it's the assessment of the British government's security agency. Meanwhile, in California, workers at major AI labs have begun organizing. Unionizing AI isn't a joke — it's the natural response to an industry that burns through talent at an unsustainable pace.

And to close the circle: $160 million in grants for AI startups launched by the UK. The British government knows AI is strategic and is putting money where others put speeches.

Three realities that feed into each other: a model anyone can use, governments no longer ignoring the risks, and workers organizing. It's not an orderly industry — it's an ecosystem defining its rules on the fly. And in that chaos, having a model like DeepSeek V4 Flash within anyone's reach is the best news.

— Max