Step 3.5 Flash Undercuts Qwen Serve Costs

Hot take: total parameters are a bad proxy for API price; the bill is mostly about what has to stay hot per token and how efficiently the provider can run it.

• –Step 3.5 Flash is a sparse MoE model with 196B total parameters but only 11B active per token, and StepFun describes it as designed around inference cost and speed.
• –Qwen3.6-35B-A3B has 35B total parameters and 3B activated per token, but its serving stack includes a 256-expert MoE design, a vision encoder, and a long 256K native context, which affects deployment economics.
• –DeepInfra’s listed pricing is $0.10 input / $0.30 output per 1M tokens for Step 3.5 Flash versus $0.19 input / $1.00 output for Qwen3.6-35B-A3B, so the provider is clearly pricing for more than parameter count alone.
• –The output-token gap is the bigger tell: vendors usually charge more where decode-time throughput, long-context KV cache pressure, and demand are harsher.
• –In plain English, “4x bigger” by headline size does not mean 4x more expensive to serve; sparse activation can flip that intuition.