← sources

Was my R9700 Worth the price?

source: reddit (reddit:LocalLLaMA)published 2026-10-14 21:45ingested 2026-10-15 02:42

retained text

I have been working on a side project the past week. I am running a single R9700, Qwen3.6-35B-A3B-UD-Q6\_K with a 220k context size. I use the claude cli as my harness and I have been testing out the AWS AIDLC 2.11 workflow. 

Stats from Claude :  
Sessions: 46                    Longest session: 13h 37m 31s

Active days: 5/7                Longest streak: 3 days

Most active day: Oct 14         Current streak: 3 days

Input 20.7m · Output 3.2m · Cache read 1.4b · Cache write 0

  
A few notes... Qwen 3.60 35B has not been perfect. As context grows, it sometimes get confused with the workflow commands. There are some lost tokens, and multiple turn code reviews, until it got things perfect. Opus would probably have been better there, so the numbers below are probably not exact. On the other hand, my llama cache has no set expiration if I pause work. I have been considering paying for Opus just for code reviews and planning, I have been using ChatGPT $20 plan for that by dragging files and copying and pasting stuff i want a second agent to review, which is annoying but it works.

Estimated costs from chatgpt compared to Opus 5.5 : 

# $426.80

46 sessions · 5 active days · Standard API pricing

# Cost breakdown

|Token category|Usage|Rate / 1M|Cost|
|:-|:-|:-|:-|
|Input|20.7M|$4.00|$82.80|
|Output|3.2M|$20.00|$64.00|
|Cache reads|1.4B|$0.20|$280.00|
|Cache writes|0|$5.00|$0.00|
|Total|||$426.80|

Pricing verified against [Anthropic's official pricing documentation](https://platform.claude.com/docs/en/about-claude/pricing).

Claude Platform Docs

\+1



# Session statistics

Average per session

# $9.28

Per active day

# $85.36

Cache read share

# 65.6%

Where the $426.80 goes

Cache reads

Input

Output

# The interesting part: your cache usage

You've accumulated 1.4 billion cached input tokens. At the standard, uncached Opus 5.5 input rate, those tokens would theoretically cost $5,600 instead of $280.

Your local Qwen setup is therefore processing a workload that would have a meaningful API cost, even with prompt caching.

The $426.80 is an estimate, not a billing statement. It assumes equivalent Opus token counts, standard API rates, and no additional tool or service charges. Your displayed statistics are also rounded.

For context, you're averaging roughly $9.28 in equivalent Opus API usage per session. That is a useful benchmark when evaluating the economics of your local AMD R9700 system and the lighter workflow we're designing.

document fetched 2026-10-15 00:45 · canonical https://reddit.com/r/LocalLLaMA/comments/1x64267/was_my_r9700_worth_the_price