What's this about?
Chinese lab DeepSeek released V4-Pro-0813, a 1.6-trillion-parameter model with 49 billion parameters active at a time. Its benchmark scores land around Claude Opus 4.8 — beating it on some measures, below GPT-5.6 Sol and Fable 5 overall — but it costs roughly a twentieth as much. Official pricing is ¥3 per million input tokens on a cache miss, ¥0.025 on a cache hit, and ¥6 per million output tokens. Confusingly, DeepSeek is also raising its prices starting today.
What exactly does "1.6T params, 49B active" mean?
A model's "parameters" are the numbers it learned during training — more parameters generally means more capacity to store knowledge. But in a Mixture-of-Experts (MoE) design, the model is split into many specialist sub-networks, and only a few are switched on for any given word. So this model *holds* 1.6 trillion parameters but only *runs* 49 billion of them per token. The analogy: a hospital with 1,600 specialists on staff, but your particular visit only involves 49 of them — you get the depth of a huge institution at the cost of a small consultation.
What's the community saying?
- The price gap is the whole conversation. The top comment summarized it as "competitive with Opus 4.8 but weaker than Sol or Fable — about 20x cheaper," and much of the thread worked through when that trade is obviously correct.
- "Good enough" is winning for most tasks. A widely-agreed comment argued that Opus 5 at max effort is more intelligent than any of us, but 99.8% of real work doesn't need that — people are mixing Kimi-K3, GLM-5.2, and Minimax by cost tier rather than defaulting to the best model.
- Momentum beats capability in adoption. Several noted the puzzle that Chinese models are surprisingly capable yet developers still default to whatever became the "industry standard," suggesting habit and procurement inertia matter more than benchmarks.
- Data-residency objections. A recurring dissent: some commenters said they'd stay on Western models specifically to avoid sending data to a Chinese provider, even at a worse price-performance ratio.
- The V4-Flash comparison. Multiple people said the earlier, cheaper V4-Flash-0731 has been the standout model of recent months and they're not sure Pro is worth the step up for their workloads.