In partnership with
KNOWLEDGE LOOP · TECH INTELLIGENCE
MODEL LAUNCH · AI AGENTS
DeepSeek's V4 Pro Just Launched. Then Came The Price Hike
Stronger agents, a 1-million-token memory, and a pricing reset that lands before most developers finish reading the changelog.
6 MIN READ · AUGUST 16, 2026
Hey readers,
DeepSeek spent four months previewing a better model. On Thursday, it delivered one — then, almost as an afterthought, said the free lunch is ending too.
That second part is the real story. The model upgrade is the headline. The pricing reset, which takes effect today, is the plot twist most coverage buried three paragraphs down.
ADVERTISEMENT
Your Brand, Front and Center
The market intelligence most desks are missing.
CEOs, politicians, regulators, and analysts talk for hours on podcasts every day.
Particle actively transcribes 125,000+ podcasts and turns them into structured data, available over REST API or MCP, searchable within minutes of airing.
Track a company or person by name, with each mention cited to the exact quote, speaker, and timestamp, or search for commentary on a thesis or sector with semantic search.
Particle is built by former Twitter and Tesla engineers, using an AI-native transcription pipeline that delivers high accuracy and extensive data enrichment.
↳ Back to the Story
The Model Got Better
DeepSeek-V4-Pro-0813 is now live across the app, web, and API, closing out a preview window that opened in April. Unlike the last release cycle, this one was built around a specific bet: agents, not chat.
Under the hood, it's a Mixture-of-Experts model — 1.6 trillion total parameters, only 49 billion active per token, which is how it stays fast despite its size. It handles a 1-million-token context window and can generate up to 384,000 tokens per response, in thinking or non-thinking mode.
DeepSeek also shipped DeepSeek Harness v0.1, an open-source, MIT-licensed framework for building software agents, alongside native OpenAI Responses API support and a one-click Codex setup. On DeepSeek's own benchmarks, V4 Pro scored 87.9 on Terminal-Bench 2.1, 62.7 on DeepSWE, and 61.5 on NL2Repo — vendor numbers worth treating as a starting point, not gospel, until independent evals catch up.
1M
token context window
384K
max output tokens
1.6T / 49B
total / active params
DeepSeek V4 Pro's architecture and vendor-reported agent benchmark scores at GA
The Price Story
Here's where the launch turns into a warning label.
Through this week, V4 Pro billed at a flat $0.435 per million input tokens, $0.87 per million output, and $0.003625 per million on a cache hit. At 16:00 UTC today, that flat rate disappears for peak and off-peak pricing — and "off-peak" means half of the new peak rate, not half of the old one.
Rate | Old (flat) | New off-peak | New peak |
|---|---|---|---|
Output / 1M tokens | $0.87 | ≈ $1.98 | $3.96 |
Cache-miss input / 1M | $0.435 | ≈ $0.66 | ≈ $1.32 |
Cache-hit input / 1M | $0.003625 | higher | higher |
Peak hours run 01:00–04:00 and 06:00–10:00 UTC; everything else is off-peak. Even at the new peak rate, DeepSeek still undercuts most frontier competitors on a per-token basis — this is a repricing, not a reversal of its cost advantage.
Launch timeline: V4 previewed April 24 → V4-Flash went GA July 31 → V4-Pro-0813 went GA August 12–13 → new pricing lands August 16, 16:00 UTC — the same day this issue lands in your inbox.
Old flat rate versus the new peak and off-peak output pricing, effective today
Why Developers Should Care
For casual chat use, this barely registers. For anyone running an agent on a loop, it changes the math entirely.
Picture a coding agent that reads a file, calls a tool, writes code, checks the result, and tries again. Every retry burns output tokens — and at the new peak rate, a sloppy five-attempt loop costs roughly four and a half times what it did last week.
The shift: stop optimizing for cost per million tokens, start optimizing for cost per completed task. A cheap model that needs five tries isn't cheaper than a pricier one that nails it in two.
Cheap frontier AI may have been a customer-acquisition strategy, not a permanent cost structure — and V4 Pro is the clearest evidence yet that the entire category is about to reprice around it.
What To Try First
Long-context document analysis
Autonomous coding tasks through DeepSeek Harness
Tool-calling and multi-step research agents
Large codebase debugging via the Claude Code or OpenCode integration
DeepSeek's API is Anthropic-compatible, so pointing Claude Code or OpenCode at api.deepseek.com/anthropic is documented, not a workaround. Test it against your actual workload, not just a benchmark screenshot.
Quick Answers
When does the new pricing start?
16:00 UTC on August 16, 2026 — the same day V4 Pro's price list first appeared.
Does off-peak mean half the old price?
No. Off-peak is 50% below the new peak rate, which is itself higher than the old flat rate.
Can I run V4 Pro through Claude Code?
Yes — DeepSeek's Anthropic-compatible endpoint is an officially documented integration path.
The Bigger Signal
The race is shifting from "who has the smartest chatbot" to "who can run useful autonomous work at economics that hold up." That's a harder contest, and a more interesting one to watch.
If DeepSeek holds onto strong reasoning, real context length, and agent tooling while staying cheaper than the alternatives, it remains one of the most important options for anyone building the next wave of AI software. If the price keeps climbing every few months, that story changes fast.
Would you route production traffic through DeepSeek V4 Pro today, or wait for the pricing to settle? Hit reply — I read every one.
Talk soon,