DeepSeek V4 Pro left preview, then its pricing changed four days later — what that teaches buyers

On 12 August 2026, DeepSeek's flagship model — V4 Pro — left preview and went to general availability. Four days later, on 16 August, its pricing structure changed in a way that's worth using as a case study, because it's a pattern more labs are likely to follow.

What actually shipped on 12 August

V4 Pro handles a 1M-token context window, can produce up to 384,000 tokens in a single response, and runs in either a thinking or non-thinking mode depending on the task. On agent-focused evaluations it scored 87.9 on Terminal Bench 2.1, 62.7 on DeepSWE and 61.5 on NL2Repo — a genuine signal that this release was built and marketed for tool-using, multi-step agent work, not chat.

The pricing change four days later

From 16 August 2026, DeepSeek moved V4 Pro to peak/off-peak API pricing: output tokens during peak hours cost double the off-peak rate — up to $3.96 per million at peak, against a prior flat rate around $0.87. That is not a price increase in the conventional sense — off-peak users may see little change — but it is a structural shift, and it landed within days of general availability, not as a slow-rolled, well-telegraphed change.

Why this matters beyond DeepSeek

A launch-day price is frequently the least durable number in any model announcement — this session's own reporting on Claude Sonnet 5 and o3 found two more examples of post-launch price moves in opposite directions, on different labs, within the same eleven-week window. DeepSeek's move is notable specifically because of how fast it came and because peak/off-peak tiering is a genuinely different pricing SHAPE, not just a different number — it changes how a team should architect scheduling for batch or non-interactive workloads to land in the cheaper window.

The buyer takeaway

Don't lock a production cost model to a launch-day rate card, and don't treat "flat per-token pricing" as a permanent property of a vendor's API — DeepSeek's V4 Pro shows both assumptions failing within the same two-week window. For any workload where API cost is a meaningful line item, build in a recurring check of the vendor's current pricing page rather than a one-time number captured at integration time, and where a workload can tolerate scheduling flexibility, design for a peak/off-peak split as a default assumption rather than an edge case.

---

Sourced from Unite.AI's coverage of the V4 Pro GA release and pricing change. DeepSeek V4 Pro launch recorded on TrustList's board, 2026-08-12.