DeepSeek V4: What's Actually Happening on July 24 (It's Not What Most Headlines Say)

Meta description: DeepSeek's old API model names stop working today, July 24. Here's what's actually happening, what V4 really is, and how to migrate before 15:59 UTC. DeepSeek V4: What's Actually Happening on July 24 (It's Not What Most Headlines Say) I keep seeing "DeepSeek V4 launches today" posts, and I want to save you the confusion before you build a whole workflow around the wrong date. DeepSeek V4 didn't launch today. It launched three months ago, quietly, as an open-weight preview — and most of the internet missed it because it didn't come with the fanfare of DeepSeek's R1 moment back in January 2025. What's actually happening today, July 24, at 15:59 UTC, is a deadline: the old model names powering a huge number of production apps stop working, permanently, unless they've already been migrated. The direct answer: DeepSeek V4 (Pro and Flash variants) has been publicly available since April 24, 2026. Today's actual news is that the legacy API model names deepseek-chat and deepseek-reasoner are being fully retired at 15:59 UTC, forcing every developer still using them to switch to the explicit deepseek-v4-pro or deepseek-v4-flash model IDs — or lose API access entirely. Quick Facts Detail Info V4 Preview launch April 24, 2026 (confirmed by Reuters) V4 GA rollout Reportedly began around July 19, 2026 Legacy API retirement deadline Today — July 24, 2026, 15:59 UTC Models affected deepseek-chat and deepseek-reasoner aliases What to switch to deepseek-v4-flash (fast/cheap) or deepseek-v4-pro (stronger reasoning) License MIT — open weights, commercial use permitted V4-Pro size 1.6 trillion total parameters, ~49B active (Mixture-of-Experts) V4-Flash size 284 billion total parameters, ~13B active Context window 1 million tokens (both variants) New pricing structure Peak/off-peak ("peak-valley") pricing — peak hours cost roughly 2x off-peak What's Actually Happening Today Here's the part that matters if you have anything in production talking to DeepSeek's API: the model aliases deepseek-chat and deepseek-reasoner — the names most tutorials, scripts, and integrations have used since DeepSeek first got popular — stop resolving to anything after 15:59 UTC today. During the preview period, those aliases were quietly routed behind the scenes to deepseek-v4-flash , so most people didn't notice anything changed. After today, that routing stops, and any code still calling the old names by name will simply fail. The fix, if you haven't already made it, is almost insultingly simple: open your integration code, change the model field from deepseek-chat or deepseek-reasoner to deepseek-v4-flash or deepseek-v4-pro , and everything else — your base URL, your API key, your request structure — stays the same. DeepSeek's own migration guidance frames this as a rename with a deadline, not a rebuild. If you've been putting off checking your integration code because "DeepSeek hasn't changed anything," today is the day that stops being true. The Real Timeline (Because Most Coverage Has This Backwards) Date What Happened April 24, 2026 DeepSeek V4 Preview launches publicly — V4-Pro and V4-Flash, open-weight, MIT license, confirmed by Reuters Late June 2026 DeepSeek announces new peak/off-peak API pricing to take effect alongside the eventual full release July 19, 2026 GA (General Availability) reportedly begins rolling out, per community and developer reporting July 24, 2026 (today) Legacy deepseek-chat / deepseek-reasoner aliases retire at 15:59 UTC Why this matters to you: if a headline today told you a brand-new model "just dropped," it's conflating two separate, several-months-apart events into one. The model itself isn't new. The deadline is. What DeepSeek V4 Actually Is Both V4 variants use a Mixture-of-Experts architecture, meaning the model has a very large total parameter count but only activates a fraction of it for any given request — V4-Pro activates roughly 49 billion of its 1.6 trillion total parameters, while V4-Flash activates around 13 billion of its 284 billion total. That's the same basic efficiency principle behind most frontier models released this year: bigger overall capacity, without paying the full compute cost on every single token. Both models ship with a 1-million-token context window by default and support what DeepSeek calls Thinking and Non-Thinking modes — essentially a toggle between fast, direct responses and a slower, more deliberate reasoning process for harder problems. The underlying attention mechanism reportedly combines two techniques DeepSeek calls Compressed Sparse Attention and Heavily Compressed Attention, aimed specifically at cutting serving costs rather than chasing raw benchmark supremacy. That's consistent with DeepSeek's entire strategy since R1: it has never really tried to be the single smartest model on the leaderboard. It tries to be the cheapest model that's smart enough. How V4 Actually Performs (With the Caveats That Matter) DeepSeek's own technical materials report V4-Pr