Home / Choosing your model and your tool
Local model, custom-built, or simply a better prompt
An open-weight model run locally involves no cost charged per call and sends no data outside, a custom-built model costs a lot to construct and is only worth it at very high volume or for a fixed format, and in the vast majority of cases a better-written prompt solves the problem faster than either.
An open-weight model installed on one's own machine or server involves no cost charged per call, unlike a paid API, and no data leaves for an outside service since everything runs locally. A custom-built model, trained specifically for a fixed format or vocabulary, costs a lot to construct. Between the two, the question to ask first remains almost always the same: would a better-written prompt already be enough.
The local model, what it actually changes
No charge per call does not mean no cost at all: it takes hardware capable of running it, or a service rented by the hour, and electricity consumed during each computation. What the local model really brings is control: no data leaves the machine or the internal network, which matters when the text handled contains sensitive information that an internal policy forbids sending to an outside service. That control has a price, the price of hardware and maintenance, and it should not be chosen for the sole reason that an API call costs something per use.
The custom-built model, a heavy undertaking
A custom-built model takes an existing model and retrains it on a dataset specific to a repeated task, to fix a format or a tone once and for all. Claude's API does not currently offer this in direct form, according to Anthropic's official glossary consulted on 2 September 2026, a situation that can change and that should be rechecked before any decision of this kind. A custom fine-tuning, where a provider offers it, is only worth it at very high repeated volume or for a format that stays fixed: building this undertaking for a need that is still evolving amounts to pouring concrete over an idea that has not yet settled.
The reflex to try first
In the vast majority of cases a small organisation runs into, the problem one thinks needs a local or custom model is solved faster by a better-written prompt: a more precise role, an added example, an imposed format.
Initial role: you are an assistant.
Rewritten role: you are the senior accountant of an agricultural cooperative, you know French VAT rules and you explain the reason behind every line.The lesson on choosing a model by task rather than by brand already sets out this starting reflex, an order to follow before considering a heavier undertaking. Taking a need one thought was technical and checking first whether a prompt is enough avoids an investment that nothing yet justified.
Cost and delay, heavy model against rewritten prompt
No cost charged per call for a local model, but hardware to buy or rent and electricity consumed; a custom-built model costs a lot to train and is only worth it at very high repeated volume.
No hardware to buy, no training to run: an adjustment to the wording of the request that is tested in a few minutes and corrected just as fast if the result does not fit.
A cooperative manager has to reformat an identical dashboard every month, with the same columns and the same vocabulary. He starts comparing the price of a custom fine-tuning service offered by an outside provider for this recurring format. Before signing, he first rewrites the prompt he already uses with a complete example of the expected format, then reruns the same task with this modified prompt.
Write in one sentence what this situation establishes, and in one sentence what it does not establish.
What this establishes: The situation establishes that a prompt rewritten with a complete example of the expected format was tested on the same recurring task before any decision to buy a custom fine-tuning service.
What this does not establish: It does not establish that this rewritten prompt produces a result as reliable as the custom fine-tuning on offer, nor how many times it would need to be checked before being considered stable.
The three most common miscalibrations
- Trop large This situation shows that a better-written prompt always replaces a custom fine-tuning service, whatever the task concerned.
- Trop étroit This situation shows nothing usable, since it concerns a single month of dashboard tested once.
- À côté This situation shows that the outside provider charges too high a price for this type of recurring format.
- An open-weight model installed locally involves no cost charged per call, but it requires hardware to buy or rent and electricity consumed.
- A custom-built model costs a lot to construct and is only worth it at very high repeated volume or for a format that stays fixed.
- Claude's API does not currently offer direct fine-tuning, according to Anthropic's official glossary, a dated situation that should be rechecked before any decision.
- In the vast majority of cases, a better-written prompt, with a more precise role or an added example, solves the problem faster than a local or custom-built model.
- Checking first whether a prompt is enough avoids launching a local or custom-model undertaking for a need that does not yet need it.
Take a repetitive task you have Claude carry out every week or every month, add a complete example of the expected format at the start of your usual prompt, and rerun this task once to compare the result obtained against the one from before.
These points depend on an interface or a rule that may have changed since this was written. Check them on your own screen before relying on them.
- The status of direct fine-tuning on Claude's API may change after this lesson was written: check the official glossary page before basing a purchase or non-purchase decision on it.
Every datable claim in this lesson links here to the public text behind it. A source that does not open proves nothing.
- Anthropic, glossary, status of fine-tuning on Claude's API consultée le 2026-09-02