DeepSeek's V4 is the loudest open-weight release of the summer, and it did not arrive quietly. The lab shipped two models, flipped its entire API over to them, and gave its old workhorses a hard shutdown date. If you were still calling deepseek-chat or deepseek-reasoner, those endpoints went dark after July 24, 2026 — everything now routes to V4.
Here is what actually changed, and why it matters beyond the launch-day noise.
Two models, one very large family
V4 ships in two flavors, both mixture-of-experts, both open-weight and published on Hugging Face under the deepseek-ai/DeepSeek-V4 collection.
| Model | Total params | Active params | Context | Max output |
|---|---|---|---|---|
| DeepSeek-V4-Pro | 1.6T | 49B | 1M | 384K |
| DeepSeek-V4-Flash | 284B | 13B | 1M | 384K |
The headline number — 1.6 trillion total parameters — sounds enormous, but the point of a sparse MoE is that you never pay for all of it. V4-Pro activates only 49B parameters per token, so inference cost tracks the active count, not the total. Flash goes further, activating just 13B of its 284B. That is the whole trick behind the pricing, which we'll get to.
Both models default to a thinking mode and can be switched to a faster non-thinking mode, and both speak the OpenAI Chat Completions and Anthropic API formats. That dual compatibility is deliberate: DeepSeek explicitly integrated V4 with Claude Code, OpenClaw, and OpenCode, so you can drop it into an agentic coding harness without rewriting your client.
The real innovation is attention, not parameter count
The parameter numbers grab attention, but the more interesting engineering is how V4 handles long context. DeepSeek describes its approach as token-wise compression combined with DSA — DeepSeek Sparse Attention.
"Novel Attention: Token-wise compression + DSA (DeepSeek Sparse Attention). World-leading long context with drastically reduced compute and memory costs." — DeepSeek's official release notes
The practical result is that 1M-token context is now the default across every official DeepSeek service, not a premium tier you opt into. Sparse attention is what makes a million-token window economically sane: instead of every token attending to every other token, the model learns which connections actually matter and skips the rest. That is the difference between a 1M window being a marketing bullet and a 1M window being something you can afford to run.
Pricing is the story
Open weights are nice. The API pricing is what will move usage. Here is what DeepSeek charges per 1M tokens:
| Input (cache hit) | Input (cache miss) | Output | |
|---|---|---|---|
| V4-Flash | $0.0028 | $0.14 | $0.28 |
| V4-Pro | $0.003625 | $0.435 | $0.87 |
Read those cache-hit input prices again. V4-Flash charges $0.0028 per million cached input tokens — fractions of a cent. Context caching turns repeated-prompt workloads (agents re-reading the same codebase, RAG pipelines hitting the same documents) into a rounding error on your bill.
Even at full cache-miss rates, V4-Pro's $0.87 per million output tokens undercuts every Western frontier model by a wide margin. For teams that have been watching their Claude or GPT bills climb with every agent loop, that gap is the entire pitch.
Does it actually perform?
Cheap is only interesting if the model is good. DeepSeek's own positioning is measured rather than boastful: V4-Pro leads all current open models on world knowledge and trails only Gemini-3.1-Pro, while claiming open-source state-of-the-art on agentic coding benchmarks.
Independent write-ups peg V4-Pro at roughly 80.6% on SWE-bench Verified — the benchmark that measures real GitHub issue resolution — putting it within a fraction of a point of the top closed models rather than in a separate tier. On harder cross-domain reasoning like Humanity's Last Exam, it lands lower, in the high-30s percent, behind the frontier closed models. So the honest read is: frontier-adjacent on coding, a step behind on the hardest reasoning, and priced like neither.
That profile is exactly what you want for the workload V4 is aimed at. Agentic coding is mostly tool calls, file reads, and iterative edits — a regime where an 80%-on-SWE-bench model at a tenth of the price beats a 81% model you can barely afford to loop.
What to actually do about it
If you run agents or RAG, the migration math is simple. Point your existing OpenAI- or Anthropic-format client at https://api.deepseek.com, change the model string to deepseek-v4-pro or deepseek-v4-flash, and keep your base URL. The concurrency ceilings differ — 2,500 for Flash, 500 for Pro — so batch-heavy workloads will want Flash anyway.
If you self-host, the weights are on Hugging Face today. A 1.6T-parameter model is not a single-GPU affair, but the 284B Flash variant is far more approachable, and its 13B active count means inference throughput punches well above its total size.
One caution: the old endpoints are gone. If you had deepseek-chat or deepseek-reasoner hardcoded anywhere — a cron job, a side project, a forgotten Zapier — it stopped working after July 24. Grep your configs.
The Bottom Line
DeepSeek V4 is not trying to win the "smartest model" crown, and it says so. It is trying to make frontier-adjacent capability boringly cheap and openly available, with 1M context as the default and a sparse-attention design that makes that window affordable. For anyone building agents or long-context pipelines on a budget, V4-Flash at $0.0028 per million cached tokens is the kind of number that quietly rewrites what's possible. The frontier labs still hold the top of the leaderboard — but DeepSeek just moved the floor a lot higher, and gave the weights away.
