DeepSeek API Pricing Update

Original Article ↗ Hacker News Discussion ↗

DeepSeek has sharply increased API prices for its V4 Flash and Pro models, especially for cached tokens and during new peak-hour windows, eroding its previous position as an ultra‑cheap option. Commenters weigh how the higher, time‑of‑day–dependent rates compare with rivals like OpenAI’s Luna and other Chinese models, noting that DeepSeek remains relatively affordable but far less of a “practically free” outlier. The changes raise questions about capacity constraints, the future of low-cost AI access (particularly in lower‑income regions), and whether third‑party providers or subscriptions will now offer better value.

Pricing Changes and Magnitude

  • Users report ~2–3x increases overall, with much steeper hikes on cache hits, especially for DeepSeek V4 Pro.
  • Shared tables show:
    • Flash off‑peak: ~1.5–2.5x increase; peak: ~3–5x.
    • Pro off‑peak: ~1.5–6x; peak: ~3–12x, with cache hits seeing the largest multipliers.
  • Peak pricing now makes Flash output markedly more expensive than some OpenRouter providers; off‑peak is about half that.
  • Some note DeepSeek is still cheaper than OpenAI/Anthropic frontier models on a pure token basis, especially for cache reads.

Impact on Use Cases (Especially Agentic Coding)

  • Heavy agentic / coding workflows rely on cache hits for 90%+ of tokens; the 6–12x cache hike on Pro is seen as particularly painful.
  • One camp argues Pro is now “bad value” versus Flash and close competitors, especially compared to subscription plans elsewhere.
  • Another camp insists even the new prices are “still basically free” at human time scales and DeepSeek remains cost‑effective.

Competition and Alternatives

  • Many compare directly with GPT‑5.6 Luna:
    • Before: DeepSeek much cheaper.
    • After: cost closer; some say Luna is now the better deal, especially given speed.
  • Multiple third‑party providers (OpenRouter, Opencode, Baseten, token aggregators) serve DeepSeek and other models.
  • However, commenters stress that alternative DeepSeek hosts generally had much worse cache‑hit pricing even before this change, so “old DeepSeek prices” were never really matched elsewhere.
  • Some expect other providers to raise prices; others note no one had matched DeepSeek’s cache pricing anyway.

Peak/Off‑Peak, Geography, and Demand Management

  • Peak hours (01:00–04:00, 06:00–10:00 UTC) line up with Chinese work hours and Asian mornings, suggesting demand is mostly domestic/Asian.
  • US and European daytime often falls into off‑peak, which some see as a net win. Night‑owl users in the US are hit harder.
  • Many interpret the change as capacity management rather than pure margin grabbing, with hopes prices could fall if capacity improves.

Open Weights, Privacy, and Policy Constraints

  • DeepSeek’s open weights and multiple vendors are seen as a major advantage: self‑hosting, redundancy, and bargaining power.
  • At the same time, DeepSeek’s official API privacy policy (broad rights to use conversation data) is criticized as unusually permissive.
  • Some are more willing to accept this because the resulting models are released as open weights; others see it as an unacceptable risk.
  • Several note that many Western companies either forbid or are wary of using Chinese models, regardless of technical merit.

User Sentiment and Affordability

  • Light users see personal bills going from a few dollars to low tens per month; most still find this acceptable.
  • Others emphasize the hit to users in low‑purchasing‑power countries, for whom DeepSeek had just become truly accessible.
  • There is visible fatigue with rapid pricing and product churn across the AI ecosystem, though some argue this is normal in a fast‑moving, early market.