HN Debrief

DeepSeek API Pricing Update

  • AI
  • Infrastructure
  • Developer Tools
  • Open Source

DeepSeek posted a new API pricing table for V4 Flash and V4 Pro that raises input and output token prices by roughly 1.5x to 5x depending on model and whether usage lands in off-peak or peak windows. The biggest shock is cache-read pricing. V4 Pro cache reads jump more than 6x off-peak and more than 12x at peak. That matters because coding agents and long-lived sessions often reuse the same context over and over, so cache hits can dominate total token volume.

If you built around DeepSeek’s old economics, rerun your cost model now using real cache-hit rates and time-of-day usage. The open-weights story still gives you fallback options, but the cheapest advertised provider is not necessarily the cheapest for long agent sessions.

Discussion mood

Mostly disappointed but not shocked. People liked DeepSeek because it made agentic coding almost comically cheap, and this update kills that feeling, especially on cache-heavy Pro usage. Still, many accept the increase as a straightforward capacity response rather than a bait-and-switch, and several note it remains competitive against major closed-model APIs.

Key insights

  1. 01

    Cache pricing breaks old coding economics

    Cache-read costs were the hidden engine behind DeepSeek’s popularity for coding agents. Long sessions can be 90 percent or more cache hits, so the jump in cache pricing hits much harder than the headline input and output numbers suggest. That is why V4 Pro suddenly looks weak on value even if its base token rates still appear competitive on paper.

    Audit your agent workloads by cache-hit share before choosing a model. If your sessions are long and iterative, cache pricing now deserves its own line item in vendor selection.

      Attribution:
    • petercooper #1
    • benjiro29 #1
    • Tiberium #1
  2. 02

    This looks like capacity rationing

    The more convincing explanation is not that DeepSeek mispriced its product forever, but that demand outran available compute. Comments cite independently reproduced old prices and a translated CEO transcript describing limited Huawei capacity and a sizable performance gap versus NVIDIA hardware. That makes time-based pricing look like a blunt tool to keep the service usable while supply stays constrained.

    Treat low API prices from fast-growing model vendors as temporary unless you know their compute position. If a vendor is supply constrained, expect pricing and availability to move together.

      Attribution:
    • dhx #1 #2
  3. 03

    Effective cost depends on cache hit quality

    A cheap posted cache rate is only valuable if the provider actually delivers cache hits. Comments from OpenRouter users say routing across providers led to misses and reliability problems, which pushed real costs above expectations until they pinned DeepSeek as the preferred backend. That turns pricing into an operational issue, not just a rate-card comparison.

    Measure actual cache-hit rate and routing behavior in production instead of trusting aggregator defaults. A slightly pricier provider can be cheaper overall if it preserves cache locality.

      Attribution:
    • lcampbell #1
    • skeledrew #1
    • nyargh #1
  4. 04

    Open weights still buy customer leverage

    Even after the increase, DeepSeek keeps one advantage Luna and similar closed APIs do not. You can self-host it or buy it from multiple inference vendors. That does not guarantee the old bargain pricing, but it does give buyers fallback paths, more negotiating power, and stronger privacy options than a single-vendor model can offer.

    If model portability and vendor leverage matter to you, keep open-weights models in your stack even when their first-party API stops being the absolute cheapest. The option value becomes obvious when pricing changes overnight.

      Attribution:
    • culi #1
    • eli #1
    • cmrdporcupine #1

Against the grain

  1. 01

    Luna may now be the simpler choice

    For buyers who only care about price and latency, the new table strips away much of DeepSeek’s old appeal. The argument here is straightforward. If Luna is now around the same price for the relevant workload and returns answers about twice as fast, the open-weights upside may not justify the operational complexity.

    Re-benchmark end-to-end task latency, not just per-token pricing. If your product is user-facing, faster completion can outweigh a small model-cost advantage.

      Attribution:
    • andai #1
  2. 02

    Peak pricing could be temporary pressure relief

    One skeptical read is that this is not a lasting repositioning at all. It is just a way to throttle demand during overloaded hours. If capacity catches up, prices could fall again, which means some teams may be overreacting if they assume the new schedule is permanent.

    Avoid hard rewrites based on a single pricing move if DeepSeek is not mission critical for you. Add monitoring and contingency plans, but wait for a few billing cycles before treating this as the new floor.

      Attribution:
    • reddec #1
    • kortzeus #1

In plain english

API ↩
Application Programming Interface, a defined way for software to expose functions or data to other software.
inference ↩
The stage where a trained AI model is used to generate outputs or make predictions.
open weights ↩
AI models released with their learned numerical parameters so others can run or adapt them.
OpenRouter ↩
A service that routes model API requests across multiple model providers.
self-host ↩
To run software or a model on your own infrastructure instead of using the vendor’s hosted service.

Reference links

Primary pricing references

DeepSeek cost and capacity analysis

Competitive benchmarks and alternatives