The cheapest LLM for each use case

The cheapest LLM depends on your task: look for low input prices when reading large amounts of text, low output prices when generating it, and a low cost per accepted result for coding or automation. In the selection below, Ling 3.0 Flash is the cheapest paid candidate per token; an offer listed at zero still needs to be assessed against its access terms. Among vendors with tight pricing, Mistral AI details its lineup and rates.

A pricing selection, not a quality ranking

The prices below come from the site’s October 2026 pricing survey, updated on October 24. They are listed in dollars per million tokens, with separate rates for content sent to the model and content it generates. These prices establish nothing about the quality of its French, the reliability of its data extraction, or its ability to edit code.

The selection covers several useful pricing patterns. It does not claim to identify the cheapest option across every provider and billing arrangement.

Candidate	Input / output	Pricing pattern to consider	What to check before choosing
Ling 3.0 Flash	$0.021 / $0.063	Low prices for both input and output	Quality on your tasks and availability of the offer
Nex-N2.5-Mini	$0.025 / $0.10	Another low-cost candidate	Instruction following and how often results need reworking
DeepSeek V4.1 Flash	$0.02 / $1	Cheap input, substantially more expensive output	The response length you are actually billed for
Flint	$0 / $0	Zero price in the survey	Quotas, terms, and continued access

Source for every price in the table: the site’s October 2026 pricing survey, available through the linked model profiles. No particular capability is inferred from these models’ names.

For more options, browse the language model profiles. The LLM benchmark lab offers another perspective to weigh against your own examples: price and capability remain separate criteria.

1. Classifying messages and extracting fields

To assign a category to a message, identify a date, or extract an amount, start with a candidate that has low prices for both input and output. In this selection, that makes Ling 3.0 Flash the first candidate to evaluate on price, followed by Nex-N2.5-Mini. That starting point does not establish whether either model can do the job.

Keep the expected output short and easy to validate. Define the allowed categories, the format of each field, and what should happen when information is missing. A field left empty for a valid reason may be preferable to a value filled in without supporting evidence.

Compare results against reference answers. Include incomplete messages, ambiguous dates, and documents containing several amounts: these cases expose the cost of corrections.

If a deterministic rule is enough, use it. Checking whether a field contains an expected address or searching for a known identifier does not necessarily require a language model. The lowest budget may come from eliminating a call altogether.

The useful metric is cost per correct extraction, including validation. A low rate loses its appeal if checking each result manually takes as long as entering the data yourself.

2. Summarizing long documents

Summarization generally sends more text to the model than it produces. Input pricing therefore carries more weight, but output still belongs in the calculation. DeepSeek V4.1 Flash illustrates this trade-off: its input rate is slightly lower than Ling 3.0 Flash’s in the survey, while its output costs more.

Consider this hypothetical scenario, calculated from the table’s prices: one million input tokens and ten thousand output tokens would cost $0.02163 with Ling 3.0 Flash, compared with $0.03 with DeepSeek V4.1 Flash. These amounts exclude additional charges and say nothing about summary quality.

The lowest input price is therefore not enough to win, even when the response is short. A token is a unit the model uses to process content; do not treat a page count as a direct measure of billable usage.

Pay particular attention to omissions: deadlines, exceptions, qualifications, and conflicting information. Ask for a summary supported by identifiable passages in the document, then check those references.

For a large collection of documents, select the relevant passages before making the call when the task allows it. Sending less text can save more than switching models, provided you do not remove information essential to the answer.

3. Writing, rewriting, and translating

When you generate large amounts of text, start with output pricing. Among the paid models selected here, Ling 3.0 Flash is the initial candidate on price. Its suitability for writing in French or translating still needs to be established using your own content.

A short rewrite and a technical translation require different checks. For the rewrite, assess meaning, length, and tone. For the translation, also check terminology, proper names, numbers, and negations.

Count every version generated. A cheap first draft followed by repeated revisions can cost more than a response you can use immediately. The length of the final text does not necessarily reflect all the usage you paid for.

Give precise instructions: audience, register, expected length, and terms to preserve. Asking directly for the deliverable also avoids paying for an explanation you do not need.

For documents intended for publication, include review time in the budget. Token savings matter less if they substantially increase the work needed to correct the text. The acceptable threshold depends on the use case: an internal draft, a sales sheet, or a contractual document.

4. Answering questions from a document collection

For answers based on your documents, the cheapest model is the one that remains sufficiently faithful to the passages supplied. Start with the low-cost candidates in the table, without assuming that a larger available context guarantees a better answer.

Prepare questions that the documents answer, along with questions they leave unanswered. The model needs to recognize when the information is absent. A fluent answer without supporting evidence does not count as a success.

Separate the costs: retrieving passages, any storage, generation, and validation. Replacing the model only reduces the portion of the bill attributable to that model.

If the same instructions or documents appear repeatedly, look into caching. The official prompt caching documentation, consulted for this article, describes reusing request prefixes to reduce costs and latency.

This option depends on the API and its terms. Do not automatically apply a cache discount to the standard rates in the table: check cache writes, reads, retention periods, and how much content is actually reused.

5. Coding and running automations

For code or automation, pricing alone cannot establish which model is most economical. A response can look correct while failing tests, calling the wrong tool, or triggering another attempt.

Start by assigning inexpensive candidates tasks you can verify: explaining a function, proposing a limited change, or drafting a test. Evaluate changes involving several files or a business decision separately.

Measure spending through to the accepted result: reading context, generating responses, making errors, correcting them, and checking the outcome. Set a stopping rule as well, so a stalled task does not keep accumulating calls.

Routing can send simple requests to an inexpensive model and pass failures to another model you have already evaluated. Measure whether this helps: if almost every request needs another attempt, the first call adds an expense.

The guide to reducing AI agent costs explains these approaches in more detail. For product differences, consult the existing comparisons, including the DeepSeek V4.1 Flash vs. GPT-6.1 Sol comparison, rather than inferring capabilities from prices.

6. Processing batches or running locally

If processing can wait, batch offers provide another option to compare. They may change which candidate is most economical, but they mainly suit tasks that can tolerate a later response.

The official batch processing documentation describes this asynchronous approach. The official Gemini API pricing page, consulted for this article, also distinguishes free access, paid access, and batch processing. Check the terms that apply to your chosen model before setting a budget.

A zero price, such as Flint’s in the survey, is worth trying if the terms suit your needs. It does not establish unlimited or lasting access. Plan for what happens if the service becomes unavailable or your usage grows.

Running locally follows a different cost model: hardware, electricity, maintenance, and setup time replace some or all of the usage bill. On a computer you already own, the additional cost may be low; buying a machine solely to save money requires a calculation.

Dedicated hosting is not automatically economical either. The Hugging Face pricing documentation describes billing based on the infrastructure selected. Capacity you rarely use remains a cost to examine. The article on local LLMs and their limitations helps put this option in context.

Choose based on your bill and your results

The starting calculation is straightforward: input volume multiplied by its rate, plus output volume multiplied by its rate. Then add all applicable charges and every attempt. The LLM price calculator lets you compare these assumptions.

To choose between candidates, keep a representative set of tasks and apply the same acceptance rule throughout. Record total spending, accepted results, corrections, and completion time. Divide spending by the number of successful tasks.

• Extraction and classification: start with low input and output rates, then measure errors.
• Summarization: weigh input and output according to your actual volumes.
• Writing: count revisions and review time.
• Code and agents: budget for the full process through validation.
• Free access, batches, and local models: account for access terms and associated costs before choosing.

Frequently asked questions

What is the cheapest paid LLM?

That depends on which options you compare and your ratio of input to output. Among the prices selected for this article, Ling 3.0 Flash has the lowest rates for both. This is neither a comprehensive market ranking nor a measure of quality.

Does an AI assistant subscription include API calls?

Do not assume it does: check what your plan includes and which account is billed for API calls. To compare a subscription with an API, start with the tasks each can handle, usage limits, and the features you need.

Can you set a maximum LLM budget?

Check whether your provider offers a hard spending cap or only alerts. Add a response length limit, a maximum number of attempts, and a stopping rule to your application. An estimate does not replace monitoring actual usage.

Why do providers charge different prices for the same model?

Hosting, commercial terms, and service options can differ. Compare the exact model version, rate limits, and additional charges as well. A matching model name alone does not guarantee an identical service.