Price first — DeepSeek V4 Pro is cheaper on both sides
DeepSeek V4 Pro at $1.80 input / $3.60 output per 1M tokens. A95B at $2.00 / $6.00. On a 1:3 input:output workload the blended rate is $3.15 on V4 Pro against $5.00 on A95B — V4 Pro is 37% cheaper on the same workload. The gap is 32% at 1:1 ($2.70 vs $4.00) and 39% at 1:9 ($3.42 vs $5.60) because V4 Pro's output price is 40% lower.
| Workload (input:output) | Qwen3.8-2.4T-A95B | DeepSeek V4 Pro | Saving on V4 Pro |
|---|---|---|---|
| 1:1 (chat-shaped) | $4.00 / 1M blended | $2.70 / 1M blended | 32% |
| 1:3 (agent-shaped) | $5.00 / 1M blended | $3.15 / 1M blended | 37% |
| 1:9 (long-output) | $5.60 / 1M blended | $3.42 / 1M blended | 39% |
| Input only (rerank / classify) | $2.00 / 1M | $1.80 / 1M | 10% |
| Output only (synthesis) | $6.00 / 1M | $3.60 / 1M | 40% |
Both are open-weight frontier MoE checkpoints
A95B: 2.4T total parameters, 95B active per token, 262K native context extensible to 1M via YaRN. DeepSeek V4 Pro: published in DeepSeek V4 series as the Pro tier; exact parameter count and active count not consistently reported by the vendor. Both ship under open-weight terms. Both route through OneInfer via OpenAI-compatible endpoints.
No published same-harness benchmark table
Neither vendor publishes a side-by-side table for the two models. A95B scores 86.6 on Terminal Bench 2.1, 92.6 on GPQA Diamond, 67.7 on SWE-bench Pro, 73.5 on FrontierSWE, 56.6 on DeepSWE 1.1 — all on Qwen's published harnesses. DeepSeek V4 Pro's published numbers are on DeepSeek's own eval harness and are not directly comparable.
When to pick DeepSeek V4 Pro
- Per-token cost dominates the buying decision (high-volume agentic workloads, batch offline jobs).
- Output-heavy workloads (long-form synthesis, doc generation) where the $3.60 output rate compounds.
- You do not need the explicit 262K native context — V4 Pro's published window is sufficient for your workload.
Ready to test the workflow?
When to pick A95B
- Long-context workloads (262K native or 1M extended via YaRN) are the buying constraint.
- You need the Qwen vendor benchmark record reproduced on Anthropic / OpenAI / Google frontier harnesses.
- You are running workloads where the 37% saving on V4 Pro does not justify a model switch (low volume, pilot, evaluation).
Caveats
- V4 Pro per-token rates can change without notice; record the rate and date alongside any benchmark or production decision.
- Cache-read rates for V4 Pro are not consistently published across providers; the headline rates above are uncached.
- A95B requires thinking mode on every call. V4 Pro has not been documented as requiring thinking mode.