note · 22 October 2026
October 2026: what actually runs our office at home
We run an office of AI agents. This is one week of its usage log, read straight from the database: which models did the calls, how many were local, what it cost, and what came out the other end. No benchmark scores here, only our own usage.
What we measured
The window is 15 October 00:00 UTC to 22 October 02:43 UTC: seven full calendar days plus the first hours of the 22nd (marked with * below). Source: the table where the office logs every model call (provider, model, tokens, cost). "Local" means our own Ollama servers, logged at $0. "Hosted" means everything else: OpenRouter models, including the free-tier ones, and a small amount of Anthropic.
The setup
- Razrbot: a home PC with a consumer GPU running Ollama.
- A MacBook Pro running Ollama for the 35b model when the GPU is busy.
- A Vultr VPS runs the coordinator that hands tasks to agents and logs every call.
Calls, tokens and cost by model
| provider | model | calls | input tok | output tok | cost |
|---|---|---|---|---|---|
| local | ornith-1.5:35b | 18,552 | 315.1M | 6.6M | $0.00 |
| local | ornith-1.5:9b | 3,232 | 11.8M | 1.5M | $0.00 |
| local | ornith:9b | 7,589 | 27.6M | 4.4M | $0.00 |
| openrouter | deepseek/deepseek-v4.1-flash | 73,823 | 1,682.5M | 107.7M | $161.55 |
| openrouter | minimax/minimax-m3.1-flash:free | 789 | 17.9M | 1.8M | $0.00 |
| openrouter | nvidia/nemotron-3-super-120b-a12b:free | 693 | 17.4M | 0.4M | $0.00 |
| openrouter | nvidia/nemotron-3-ultra-550b-a55b:free | 302 | 5.3M | 0.1M | $0.00 |
| anthropic | claude-haiku-5-5 | 284 | 1.7M | 0.2M | $0.36 |
| openrouter | qwen/qwen3.7-flash | 219 | 4.3M | 0.1M | $0.11 |
| openrouter | z-ai/glm-5.3 | 125 | 3.0M | 0.2M | $1.13 |
| openrouter | inclusionai/ling-3.1-flash | 81 | 1.8M | 0.3M | $0.16 |
| all models | 105,689 | 2,088.5M | 123.4M | $163.31 | |
M = million tokens. The log has a third local tag, ornith:9b, next to ornith-1.5:9b; we list it as logged. Free-tier hosted models show $0 but are still hosted, so not "home".
Local versus hosted
By calls, 29,373 of 105,689 (27.8%) were served locally and 76,316 (72.2%) by hosted models. By tokens local models handled 17% of input and 10% of output. One hosted model, deepseek/deepseek-v4.1-flash, made 70% of all calls and $161.55 of the $163.31 spend. Its calls are large: about 23k input tokens on average, against about 17k for the local 35b.
So "mostly local" is not true of our week: about a quarter of calls ran at home, and the money went almost entirely to one hosted model. Local was ahead on only one day, 20 October.
Per day
local hosted · calls per day
| day (UTC) | local | hosted | total | local share |
|---|---|---|---|---|
| Oct 15 | 4,634 | 15,060 | 19,694 | 24% |
| Oct 16 | 2,460 | 7,529 | 9,989 | 25% |
| Oct 17 | 4,213 | 14,385 | 18,598 | 23% |
| Oct 18 | 2,878 | 6,289 | 9,167 | 31% |
| Oct 19 | 4,216 | 13,758 | 17,974 | 23% |
| Oct 20 | 6,842 | 4,221 | 11,063 | 62% |
| Oct 21 | 3,532 | 12,724 | 16,256 | 22% |
| Oct 22* | 598 | 2,351 | 2,949 | 20% |
* 22 October covers only 00:00 to 02:43 UTC.
What came out
In the same window the office marked 528 tasks as done (status done, last updated inside the window). That is about 5.0 done tasks per 1,000 model calls, or roughly $0.31 of hosted spend per finished task. Caveats: a task can take many calls, some calls are retries or planning that never end in a task, and "done" is the office's own label, not an independent quality check.
What we take from it
- Local models earn their place on volume: steady calls that would otherwise add to the bill.
- The bill is set by a single hosted model with long prompts, so context size matters more to cost than call count.
- The daily mix swings a lot (from 25% to 62% local), driven by what the agents were asked to do and which machines were free.
These numbers describe our usage pattern only: one week, one office. They say nothing about how any model would do for you. For tested results see the nightly leaderboard and the cost page.
