At NeuraGrowth we publish practical guides for people who build digital products. The research, the drafting, the niche-finding, the social posts - a lot of that work runs on its own through a system we built called Orqestra. Claude sits inside it doing most of the heavy thinking.
This post is not about AI in the abstract. It is about what one model did across 10 real days of work: how many times we called it, what it cost per call, where it succeeded, and where one run took 135 times longer than it should have.
If you have ever wondered what it looks like to hand a language model a repeating job rather than a one-off question, this is a concrete answer.
What Claude is actually doing inside Orqestra
Orqestra runs a list of operations on a schedule. On 6 September it started 23 of them. Of those, 22 finished and 15 succeeded - a 68.2 percent success rate on that single day, across operations like writing blog content, posting to social platforms, researching niches and cleaning up old product records.
Claude handles the jobs that need language. Over the 10 days from 28 August to 6 September, we called claude-sonnet-4-6 107 times across 7 different jobs: content_blog_rewrite, content_linkedin, content_note_rewrite, content_tool, content_tool_rewrite, deep_dive.research, and discover_niches.research.
That is not one task repeated 107 times. It is seven different kinds of work, each of which would otherwise require a person to sit down and do it.
The numbers: cost, speed, and zero failures
Across those 107 calls, Claude came to $7.33 in total, which works out to roughly $0.07 per call. The median response time was 11.2 seconds. The slowest single call took 499.7 seconds, which is a long wait but a separate question from the timing problem we cover below.
The failure count was zero. Not one of those 107 calls came back broken. For a pipeline that runs unattended, that matters more than the per-call price.
The most expensive single operation over those 10 days was discover_niches.research. It accounted for $5.46 of the $12.64 we spent on all operations combined - 43.2 percent of the total, across just 10 calls. At $0.55 per call on average, it is the operation where Claude is doing the deepest work.
The run that took 783 minutes instead of 6
Here is the problem. One discover_trending_niches run, started on 30 August at 09:00, took 783.7 minutes to resolve. The median across 11 runs is 5.8 minutes. That single run was 135 times slower than normal.
The operation failed. We have the timing. We do not have a logged explanation for why that specific run behaved so differently from the other ten. The record does not say.
What we can say is that the cost side and the reliability side of Claude itself were not the issue. The model returned zero failures across all 107 calls in that window. Whatever caused the slowdown was somewhere in how that particular run was handled, not in the model responses.
What a day of unattended operation actually looks like
On 6 September, Orqestra ran by itself through 23 operations. The whole day cost $1.66. The longest individual run took 5.8 minutes.
The 12 operation types it ran included things like refreshing authentication tokens, polling external balances, circulating older content, posting to Instagram and X, and writing new content. Most of those touch Claude at some point.
A 68.2 percent success rate on a single day is not perfect, and we are not presenting it as such. Some runs did not finish or did not succeed. That is the real number, and it is more useful than a rounded-up claim.
When to reach for a heavier model
The 43.2 percent cost share that discover_niches.research carries is a useful signal. It tells us where the pipeline is spending its thinking budget. Research calls are longer and more expensive because the job demands more: evaluating a niche requires pulling together more context than rewriting a short note.
We have written separately about how we chose between Sonnet, Haiku, and Opus for different jobs. The short version is that not every call needs the same model, and matching the model to the task is how you keep a 10-day bill from running away. For the research jobs, claude-sonnet-4-6 is the current choice.
If you want to go deeper on that reasoning, the post at https://neuragrowth.co/blog/sonnet-vs-haiku-vs-opus-on-critique-flashcards-when-to-reach-for-which/ covers the tradeoffs we actually ran.
What makes this worth understanding if you do not write code
You do not need to build a pipeline to find this useful. The pattern here - calling Claude on a schedule, for a fixed list of jobs, and measuring what each one costs - is available in simpler forms through tools that do not require you to write anything.
What the numbers show is that the model itself is the stable part. Zero failures across 107 calls, $0.07 per call, 11.2 seconds median. The instability in this log came from one long-running operation whose cause we cannot state, not from the model.
That distinction matters if you are deciding whether to trust a language model with a repeating job. The question is less about the model failing to respond and more about how the surrounding system handles a run that goes long.
If you are thinking about using Claude or any language model for work that runs on a schedule, measure the cost per call before you scale the number of calls up. The $0.07 we pay per claude-sonnet-4-6 call is manageable across 107 calls, but the discover_niches.research operation at $0.55 per call shows how quickly one heavier job changes the shape of the bill. Measure early, and keep the logs - because when a run takes 783 minutes instead of 6, the logs are the only place you will find out what happened. We have written more about how we think about AI model selection at https://neuragrowth.co/blog/sonnet-vs-haiku-vs-opus-on-critique-flashcards-when-to-reach-for-which/, and if you want to understand the broader shape of what NeuraGrowth builds, https://neuragrowth.co/blog/what-we-actually-make-and-who-the-family-line-is-for/ is the right starting point.
Originally published at neuragrowth.co. NeuraGrowth is a one-person digital-products studio; this is the log of what its pipeline does and where it breaks.
Top comments (0)