DeepSeek ships V4 Pro, ending its flagship's four-month preview
DeepSeek finally took the preview label off its flagship model, and Twitch handed streamers a way out of Amazon's AI training pipeline. Two stories about the same underlying fight: who controls the models, and who controls the data.
DeepSeek has released the production version of V4 Pro, its flagship model, ending a preview that ran since late April. The build — DeepSeek-V4-Pro-0813 — went live on the company's API and on OpenRouter on August 12, carrying over the preview's specs: a 1.6-trillion-parameter mixture-of-experts architecture with 49 billion active parameters, a 1M-token context window, and a 384K-token output ceiling. What changed is the label, and what the label completes: the V4 lineup is now a clean two-track offering, with Flash as the cheap, high-concurrency workhorse (2,500 concurrent requests) and Pro as the premium reasoning tier (capped at 500), priced at roughly three times Flash — $0.435 per million input tokens and $0.87 per million output — with the agent-developer tooling now expected of a flagship: structured JSON output, tool calls, a Responses API, and Anthropic-format compatibility.
On the vendor-reported numbers, V4 Pro lands right at the edge of the frontier tier. DeepSeek's model card, at maximum reasoning effort, claims 80.6 on SWE-bench Verified — level with Gemini-3.1-Pro and a fraction behind Claude Opus 4.6's 80.8 — plus the top LiveCodeBench score in its comparison table (93.5) and a 3,206 Codeforces rating. It still trails the US leaders where raw reasoning depth is the test: 67.9 to GPT-5.4's 75.1 on Terminal Bench 2.0, and 37.7 to Gemini-3.1-Pro's 44.4 on Humanity's Last Exam. None of these figures has been independently replicated for the 0813 build, and DeepSeek has not said when — or whether — the new weights will land on Hugging Face; the repos still host the April preview artifacts.
The GA is the payoff of a release strategy DeepSeek has run in public since spring. Flash went official on July 31 with agent benchmarks that beat the Pro preview, making the cheap model the default for agent workloads while the flagship waited — now Pro answers, and DeepSeek's bid to stay ahead of a Chinese field shipping flagship-class models on a near-monthly cadence (Moonshot's Kimi K3, MiniMax's M2.7) is official. Two things are worth watching: a notice on DeepSeek's pricing page promises "a significant increase" in API prices, so these rates may be the last cheap ones — and every open-weight rival just got a new reference point to undercut. We covered the independent run that confirmed Flash's agent scores — Independent run confirms DeepSeek V4 Flash's 82.7% score.
Twitch streamers can now opt out of letting their content train Amazon's generative AI. A new "Training for Generative AI" toggle in account settings keeps streams, VODs, clips, chat, and channel images and text out of future training for Amazon models that generate or synthesize text, audio, images, or video — while captions, AutoMod, and recommendation features keep working. The catches: the toggle was on by default when The Verge checked (Amazon has not confirmed the default), and if you chat on another streamer's channel, their opt-out preference governs your words. It is the formal answer to the creator backlash that followed Twitch's plans to feed streams into Amazon's AI — though an opt-out buried in settings may not satisfy streamers who wanted an opt-in.
What to watch: DeepSeek's promised price increase — and whether the 0813 open weights ever ship.
Should training-data choices be opt-out or opt-in — and do streamers deserve more than a settings toggle? Tell us in the comments.
