Conversations about picking a language model almost always open with the wrong question: which model is best? The right answer is never a brand name. It is a combination of the task to be done, the cost ceiling you can live with, and the data that is not allowed to leave your organisation.

In 2026, for ordinary enterprise text work — summarising reports, classifying tickets, extracting fields from documents — the quality gap between top-tier models has narrowed enough that it no longer decides anything on its own. Four other factors decide it.

The four criteria that actually decide

• Task fit: for classification and field extraction, a small model with a precise instruction is enough; multi-step reasoning is what needs a large one.
• Cost per answer: a number that looks trivial in a pilot becomes one of your largest operating lines at daily volume.
• Persian quality: a strong score on English benchmarks guarantees nothing about clean, idiomatic Persian.
• Data path: the question is not how capable the model is, but where the data is processed and under what commitment.

A bigger model does not mean a better answer

In most enterprise processes the desired output is not creative prose; it is a defined, repeatable format. For that work a smaller model with a clear instruction and a few correct examples delivers the same result at a fraction of the cost and latency. The measure is the acceptable-answer rate on your task, not a model's place in a leaderboard.

Measure cost per outcome, not per token

Early estimates almost always come in low, because only part of the cost is counted:

• input and output on every request, plus the system instruction and context repeated on each call;
• retries — every invalid answer multiplies the cost of that case;
• human review, which stays in the loop until confidence is earned.

Persian has to be tested, not assumed

Take fifty to a hundred real samples of your own correspondence and evaluate candidates on those: orthography and spacing, fidelity to your formal register, resistance to literal translation of domain terms, handling of numbers and dates.

Cloud or local: a decision about data

This is not a binary. Cloud is the fastest start for non-sensitive data. Local deployment protects customer records and contracts, and transfers hardware and maintenance cost to you. The third path, most common in practice, is hybrid: strip identifying fields before sending and keep source data inside the organisation.

A test that answers within a week

• Choose one specific, high-frequency task rather than a general capability.
• Build a test set from your own real data with verified correct answers.
• Evaluate two or three candidates under an identical instruction: one large model, one mid-tier, one you can host yourself.
• Record three numbers: acceptable-answer rate, cost per answer, response time.
• Pick the cheapest option that clears your quality threshold, not the highest absolute score.

Common mistakes in this decision

• Deciding from public leaderboards and a few manual prompts, with no organisational test set.
• Coupling so tightly to one provider that switching models means rewriting the system.
• Ignoring output stability; a model returning ten formats in ten runs cannot sit inside an automated process.
• Having no predefined quality threshold, which hands the comparison to whoever is in the room.

Frequently asked questions

• Should we always take the most expensive model?
  No. Reserve it for hard cases and route high-volume simple work to a cheap one; that split produces most of the saving.
• What if our data is sensitive?
  Process sensitive data locally and use cloud services for public, de-identified content.
• How often should we revisit the choice?
  Re-run the same test set every few months; the quality-to-cost ratio moves faster than a budget cycle.

Takeaway

Choosing a language model is not a technology purchase; it is an engineered decision with explicit criteria. An organisation that owns a Persian test set, has quantified its acceptance threshold and knows its cost per outcome re-decides within days whenever the market shifts. One without those criteria starts from zero every time.

Glossary

• LLM: an engine that understands and generates text; the basis of conversational assistants.
• Cost per answer: the full cost of one request at real volume, retries included.
• Test set: real organisational samples with verified answers, the basis for a fair comparison.
• Output stability: how consistent answer format and content stay across repeated runs.
• Vendor lock-in: technical dependence on one provider that makes switching expensive.