Three things about today's four model releases that the benchmark tables don't tell you

#ai #opensource #llm

Four flagship models landed within 24 hours today. The benchmark tables are getting all the attention. Here are three things I checked on the official pages that I have not seen discussed.

1. DeepSeek's price has an expiry date

The official pricing page lists $0.435 input / $0.87 output / $0.003625 cache hit per million tokens for deepseek-v4-pro.

The footnote on that same page says peak/off-peak pricing starts 2026-08-16 16:00 UTC:

	now	peak	off-peak
input	$0.435	$1.32	$0.66
output	$0.87	$3.96	$1.98
cache hit	$0.003625	$0.044	$0.022

That is 4.55x on output and 12.14x on cache hits at peak. Peak windows are 01:00-04:00 and 06:00-10:00 UTC.

If your workflow leans on prompt caching, the cache hit line is the one to budget against, not the output price.

2. "20x cheaper" is a comment, not a claim

The line everyone quotes came from a Hacker News comment posted 82 seconds after the submission. DeepSeek never said it.

In the same thread, one person computed 60x from opencode's published agentic request distribution. Another argued 10x. Straight list-price division against Anthropic's published rates gives 11.5x on input and 28.7x on output.

Three numbers, one thread, and almost nobody states whether they are comparing unit price or cost per task.

3. The open Qwen is not the Qwen in the benchmark table

Qwen/Qwen3.8-2.4T-A95B is real and downloadable: 2.4T total parameters, 95B activated, 213 safetensors shards, about 4.89TB.

But the model card says it is text-only, multimodal input is not supported, and thinking mode cannot be disabled. The vision input, non-thinking mode, default 1M context and built-in tools belong to Qwen3.8-Max, the hosted version on Qwen Cloud. That is the column in the benchmark table.

The license is also not Apache 2.0. It is a custom Qwen3.8-Max License with two commercial gates: MaaS or AI work assistant businesses with over $50M revenue in any rolling 12 months need separate authorization, and products over 100M MAU or $20M monthly revenue must display the model name prominently in the UI.