Technology

DeepSeek's New AI Model Is Here — and Using It Just Got a Lot More Expensive

Martin HollowayPublished 4month ago5 min readBased on 12 sources
Reading level
DeepSeek's New AI Model Is Here — and Using It Just Got a Lot More Expensive
source:deepseek.com

DeepReach announced DeepSeek V4 Pro on August 13, 2026, making the model available to everyone alongside a new pricing system that raises costs roughly fourfold. The new prices take effect at 16:00 UTC on August 16, according to DeepSeek's API documentation, though Reuters reported the effective date as August 17 based on a company statement.

AI companies charge for their models based on "tokens" — small pieces of text, roughly a word or fragment of a word. You pay for every 1 million tokens the model produces in response to your requests. DeepSeek V4 Pro will cost $3.96 per 1 million output tokens during peak hours, up from the current $0.87 rate. Off-peak pricing is set at $1.98 per 1 million output tokens, exactly half the peak rate (Engadget). The smaller V4 Flash model sees a comparable increase: $1.32 per 1 million output tokens at peak (up from $0.28) and $0.66 during off-peak hours. DeepSeek described the tiered pricing as a way to "allocate resources more reasonably," per the company's API documentation (DeepSeek API Docs).

The price hikes cap a confusing pricing trajectory. DeepSeek's discounted rates were originally a temporary promotion scheduled to end May 31. The company had announced it would make those promotional prices permanent but later reversed course and proceeded with the increases (Engadget). Reuters reported that the new V4 rates range from 50% to 1,100% above current prices across the model lineup (Reuters).

DeepSeek V4 Pro reaches the market as a fully open-source release, accompanied by a technical report and a live API (DeepSeek API Docs). Open-source means that the inner workings of the model — the data-driven parameters that determine how it responds — are freely available for anyone to download, study, and run on their own computers. The API is OpenAI-compatible, meaning developers already using OpenAI's software can switch over with minimal code changes. It is callable through the standard client at base URL https://api.deepseek.com and exposes a reasoning_effort parameter (accepting values such as "high") plus a "thinking" toggle passed via extra_body (DeepSeek API Docs). The flash variant has been updated to a version designated DeepSeek-V4-Flash-0731 (DeepSeek API Docs).

The V4 lineup follows a deliberate release schedule. DeepSeek-V4 Preview shipped on April 24, 2026, described as having a smaller parameter size, faster response times, and highly cost-effective API pricing with reduced compute (DeepSeek API Docs). That preview built on V3.2-Exp, released in September 2025, which had cut API prices by over 50% while improving long-context performance and reducing compute cost (DeepSeek API Docs). The arc from aggressive price cuts to fourfold increases is notable.

The new pricing repositions DeepSeek relative to competitors. Moonshot's Kimi K3 model costs $15 per 1 million output tokens. OpenAI's most advanced model, GPT-5.6 Sol, costs $30 per 1 million output tokens, while its low-cost GPT-5.6 Luna model costs $1.20 per 1 million output tokens (Engadget). At V4 Flash's peak rate of $1.32, DeepSeek now costs more than OpenAI's budget tier Luna model, a comparison that would have been implausible under the old promotional pricing.

The peak/off-peak split is a meaningful design choice. It is similar to how utilities charge less for electricity at night, when demand is lower: the company is charging more during busy hours to encourage users to shift their work to quieter periods.

For developers building on DeepSeek's API, the practical impact depends on the type of work. Batch and asynchronous tasks that can run during off-peak hours see a 50% reduction relative to peak pricing. Real-time, latency-sensitive applications bear the full peak rate. Teams currently paying $0.87 per 1 million output tokens for V4 Pro-class performance will need to budget for a 4.5x increase at peak hours starting this weekend.

The open-source release of V4 Pro, with its technical report, offers an alternative path for teams unwilling to absorb the API price increases: running the model on their own hardware. That option carries its own costs for equipment and maintenance, and the savings depend heavily on how much the hardware gets used. But the fact that the model is freely available means DeepSeek's pricing ceiling is constrained, at least in principle, by what users can achieve on their own infrastructure.

The broader context here is that the era of aggressively subsidized AI pricing may be narrowing. DeepSeek was among the most aggressive discounters in the market. If its pricing is converging upward toward competitors, the gap between budget and premium AI models may shrink rather than widen, leaving companies to compete on capability, speed, and ecosystem rather than raw price.

What remains unclear is how much of the increase is driven by underlying computing costs versus a recalibration of profit margins. DeepSeek's V3.2-Exp release in September 2025 showed that the company could simultaneously cut prices and reduce computing needs through design improvements. The V4 Preview in April was positioned similarly. The reversal to permanent low prices, followed by a quadrupling, suggests either that the computing economics shifted, that the promotional pricing was unsustainable at the volumes the company is now serving, or that DeepSeek no longer needs to buy market share through aggressive discounting.