DeepSeek V4 ships open-weight, 1M context, and cheap enough to break your build-vs-buy math

DeepSeek dropped V4 in mid-June as an open-weight model with a 1-million-token context window, priced at $1.74 per million input tokens, posting near-parity with GPT-5.4 on math and Q&A benchmarks (MindStudio). That's the headline number. The architecture underneath is more interesting. It actually shipped as two MIT-licensed Mixture-of-Experts models, DeepSeek-V4-Pro (~1.6T total / 49B active) and DeepSeek-V4-Flash (~284B / 13B active), both with 1M context and a hybrid attention mechanism built to cut inference cost on long-running agentic tasks (LLM-Stats).

Here's the part that should make you stop scrolling: V4 posts the highest reported open-weights SWE-bench Verified score at roughly 80.6%, and Flash output runs around $0.28 per million tokens. Open weights, a real coding score, and pricing that's an order of magnitude under the closed frontier coding models.

I've run the self-host-vs-API math maybe four times in the last year and it always came out the same way. Renting from a frontier lab was cheaper than running my own GPUs once you factored in ops, idle time, and the fact that the open model was a tier behind on the work I cared about. That gap just closed. Not on every axis. GPT-5.4 and Fable 5 still win on the hardest reasoning. But for agentic coding loops where you're burning tokens by the millions on a long context, an 80.6% SWE-bench model at $0.28/M output changes the calculation in a way it hasn't before.

What I'd actually do: don't rip out your stack. Run a bake-off. Take your most token-heavy agent loop, the one whose monthly bill makes you wince, and route it to V4-Flash for a week behind a flag. Measure the quality delta on your real tasks, not the leaderboard. If Flash holds 90% of the quality at 10% of the cost on your workload, that's your new default for the bulk-token path, and you reserve the expensive model for the verification step. The economics finally support a tiered routing strategy where you weren't forced into it before. That theme of routing cheap-where-you-can shows up again two stories down.