Earlier today I saw that the 17 October release of Deepseek v4 Flash is even more efficient than previous versions, threatening dominance of american AI labs.
Someone asked me about my current setup, as they burn through their weekly AI plan quickly. I did a shallow dive in the costs of the larger open weight model I use online, GLM 5.3 Flash. I didn’t use much local tokens this week, see below for my local setup this month.
I will try the latest Deepseek as soon as my current inference provider has it up and running, probably somewhere later this week.
GLM 5.3 what does it cost for a coding agent, with Zero Data Retention in Europe?
Looking at the graph, my coding agent sessions use mostly cache. A coding session tends to send the same file and tool contents over and over again, so very few new tokens are needed. Storing old tokens is relatively cheap.
The light purple part of the bars are cache, the dark purple parts are fresh tokens. Cache is relatively cheap, computing tokens costs much more, but still not that much.
I spent less than a dollar to develop Stopped-motion detection in videos, and I did some quality of life improvements on my writing environment while I was at it.
Other kinds of models can be even cheaper
I say ’Models’ in the title, but above I only mentioned Large Language Models. There is a surge in popularity of other kinds of models like Decision Models, and BERT models are making a comeback. They tend to use less compute than LLMs.
My local models in October 2026
I still use roughly the same local model setup as two months ago. QWEN3.6 27B has been replaced by QWEN3.8 27B, which is very capable in long running agentic tasks. It has less world knowledge then Deepseek V4 Flash and GLM 5.3 Flash, and is apparently more expensive to host. But I can give it a task, walk away, and it will iterate and come out with something that works usually. Walking away is sometimes necessary - my laptops fans do become noisy on long runs.
Most of my use is small things, or long running tasks where I can walk away.
My hosted open weights models at the moment
I went from Deepseek v4 Flash to GLM 5.3 flash, because the latter can process images. This has become an integral part of my workflow, e.g. for sketching out software development exercises and turning them into code. In web development it is handy if the model can control a browser, take screenshots, interpret the screenshots and iterate; synthetic exploratory testing.
I might go back to Deepseek once my provider makes the new version available, somewhere this week. It benchmarks well, so it might also work well on my code.
Buy hardware, or use a service?
People sometimes ask me what hardware they should buy for Local Models. With these prices, and my usage, this is hard to justify for pure cost reasons. If you value independence and privacy, by all means, buy some hardware. But try hosted models first to see if they work for you.