The promotional rate was supposed to end. According to DeepSeek, it isn’t.
On May 23, DeepSeek announced that the 75% discount applied to V4 Pro API access, scheduled to expire May 31, will instead become the permanent base price. Per DeepSeek’s API documentation, V4 Pro input tokens are now priced at $0.435 per million, output at $0.87 per million, and cached input at $0.0036 per million. The prior input rate was $1.74 per million tokens. That’s not a rounding adjustment. It’s a structural reset.
The catch is that the competitive significance depends entirely on what you’re comparing against. According to Artificial Analysis, an independent AI benchmarking firm that tracks cost-to-intelligence ratios across frontier models, DeepSeek V4 Pro holds a leading position on the cost-to-intelligence frontier. Artificial Analysis’s comparative benchmark data places total run costs for V4 Pro well below those of competing frontier models, figures cited from that analysis put DeepSeek’s suite run costs against several thousand dollars for GPT-5.5 and Claude Opus 4.7 alternatives, though those specific numbers haven’t been independently confirmed by TJS from the primary dashboard for this cycle. The directional picture is consistent across secondary coverage.
The pricing structure is particularly advantageous for enterprise agentic workflows that process large token volumes, where per-token costs compound significantly at scale. A workflow that runs 100 million input tokens monthly, not unusual for a moderately active multi-agent system, costs $43.50 at the new DeepSeek rate versus $174 at the prior rate. That differential widens fast against comparable frontier model pricing.
The part nobody mentions:
enterprise adoption in Western markets doesn’t run on pricing alone. Data sovereignty requirements, procurement review timelines, and geopolitical considerations around Chinese AI infrastructure can override a compelling cost argument before it reaches a budget conversation. The price is real. So is that friction.
This announcement advances a story TJS covered in May on inference cost collapse across the frontier model market. DeepSeek’s May 13 cost-performance positioning piece established the competitive framing; the May 23 permanence decision is the follow-through. Where the earlier coverage asked whether DeepSeek’s cost advantage was durable, the permanence announcement answers that question, at least on the pricing side.
Analysis
Three major frontier model pricing events in 30 days, Gemini 3.5 Flash, GPT-5.5 API, DeepSeek V4 Pro permanent, suggest the price floor across the tier-1 API market is actively contested. Enterprise teams benchmarking 2026 agentic infrastructure costs should treat these as inputs to vendor negotiations, not just adoption decisions.
What to watch
whether Anthropic, OpenAI, or Google respond with matching reductions or structural pricing changes before mid-June. Three significant pricing events in 30 days (Gemini 3.5 Flash, GPT-5.5 API, now DeepSeek permanent) suggest the frontier model pricing floor is actively contested, not settled. The next move worth tracking is whether enterprise procurement teams treat this as a switching trigger or a negotiating lever with incumbent providers.
TJS synthesis:
Don’t treat this as a discount story. DeepSeek restructured its price floor before a competitor could respond, locking in a cost-efficiency position at the moment enterprise teams are actively sizing agentic infrastructure budgets. Whether your organization can use it is a separate question from whether it changes your negotiating position with providers you can use. It does. Get the number in front of whoever owns your model spend.