DeepSeek V4 Price Doubles at Peak Hours, and deepseek-chat Dies July 24
DeepSeek's mid-July V4 launch brings the first time-of-day pricing to a frontier AI API, the legacy model names hit a hard 17-day retirement clock, and GPT-5.6's rumored July 7 launch did not happen.
DeepSeek is about to do something no other frontier AI lab has done: charge you more depending on what time it is. The official V4 launch is still scheduled for mid-July with no exact date confirmed, and when it lands, two things change at once. API bills get a peak-hour multiplier, and the model names most existing code is still calling get switched off. Here is what is confirmed on DeepSeek's own docs, what was announced, and what to do before the deadlines.
DeepSeek V4 price: the first surge pricing in an AI API
DeepSeek announced on June 30 that the official release of V4 is set for mid-July, and that it will introduce peak and off-peak API pricing for the first time. TechNode reported the Monday announcement directly. Peak hours run 9 a.m. to noon and 2 p.m. to 6 p.m. Beijing time each day, which is 1:00 to 4:00 a.m. and 6:00 to 10:00 a.m. UTC. During those windows, API usage is charged at twice the off-peak rate. Outside them, prices hold.
The rates on DeepSeek's official pricing page today become the off-peak baseline at launch. For deepseek-v4-pro that is $0.435 per million input tokens (cache miss) and $0.87 per million output tokens; at peak those double to $0.87 and $1.74. For deepseek-v4-flash it is $0.14 input and $0.28 output, doubling to $0.28 and $0.56 at peak. Both models ship a 1-million-token context window and a 384K maximum output, with concurrency caps of 500 on Pro and 2,500 on Flash. Cached input is the escape hatch: cache-hit rates sit around one percent of cache-miss rates ($0.003625 per million on Pro, $0.0028 on Flash), and caching largely sidesteps the peak multiplier because the discount is already so deep.
The direction is the surprise. DeepSeek spent the last year undercutting everyone and recently made a 75 percent V4 discount permanent, as The Next Web notes. Surge pricing pulls the other way. DeepSeek frames it as resource allocation and service stability rather than margin, the same logic Uber applies to cars, now applied to tokens. For a workload that runs around the clock and cannot shift load, Tony Reviews Things estimates roughly eight peak hours a day at 2x cost works out to a 20 to 30 percent effective price increase. For anything batchable, the play is to push async inference, evaluation runs, and background agents into the off-peak windows and keep the current rates. DeepSeek has said it will send 24-hour email notice before the pricing change takes effect, so watch your inbox for the actual cutover signal.
deepseek-chat and deepseek-reasoner retire July 24 at 15:59 UTC
The second change is a hard deadline, and it is closer than the launch. The legacy model names deepseek-chat and deepseek-reasoner will be "fully retired and inaccessible" after July 24, 2026 at 15:59 UTC, per DeepSeek's own API docs and the V4 preview announcement. That is 17 days from today. DeepSeek set this date back in April when V4 entered preview, so the deadline itself is not new, but the clock is now short enough to act on.
Both legacy names currently still work because they quietly route to V4-Flash: deepseek-chat maps to V4-Flash in non-thinking mode, and deepseek-reasoner maps to V4-Flash in thinking mode. After the cutover, a stale model string becomes a hard failure rather than a graceful redirect. The migration is deliberately small: the base URL (https://api.deepseek.com, or the Anthropic-format https://api.deepseek.com/anthropic) does not move, and you only update the model field to deepseek-v4-flash or deepseek-v4-pro. The real work is finding every place the old names live, since migration guides flag that aliases get hard-coded into environment variables, prompt routers, eval jobs, and fallback policies that are easy to miss. If you depended on deepseek-reasoner's reasoning output, test how your SDK handles the reasoning_content field on the V4 thinking modes before you swap.
Tracking
GPT-5.6: still in limited, government-gated preview. The rumored July 7 general-availability window that some prediction markets had favored did not produce a launch today, and reporting from July 5 now puts July 9 as the leading market outcome, ahead of July 14. OpenAI's only confirmed language remains "in the coming weeks" from the June 26 preview post. Treat any specific date as rumor until OpenAI announces it.
Claude Fable 5: now fully metered as of this morning's July 7 cutover. Subscription inclusion on Pro, Max, Team, and premium Enterprise plans has ended, and every Fable 5 session draws from a separate usage-credit balance at $10 per million input and $50 per million output tokens on the Anthropic pricing page.