Disclaimer

LLMTCO publishes calculators, price tables and guides on the cost of running large language models, through commercial APIs or self-hosted on GPUs. Everything on the site is general information for budgeting and comparison. It is not financial, tax, legal or procurement advice, and it is not a quote from any provider.

Prices are dated list prices

The API pricing table and the GPU pricing table show publicly listed prices with the date each one was checked and a link to the official page. API and GPU prices change often, sometimes several times a year, and your actual bill can differ because of volume or enterprise agreements, regional pricing, minimum commitments, spot or reserved capacity, currency and taxes. Every price is an editable input: confirm it on the provider's own pricing page or in a written quote before you rely on a result.

What the API cost tools assume

The token cost, monthly spend, provider comparison and cost-per-conversation tools multiply token counts by per-million-token prices, input and output priced separately, with the formula shown on each page. The caching and batch calculator applies discounts whose conditions (minimum prompt length, cache lifetime, batch turnaround) are set by each provider and may change. Rate limits, failed or retried calls, tool and image tokens and support plans are not included unless you add them.

What the self-hosting tools leave out

The GPU TCO, cloud GPU, cost-per-token, utilization and break-even tools model infrastructure cost only: hardware or rental, power and a flat overhead. They do not price engineering and operations time, latency, availability, security, compliance or data-protection work, and "cheaper" does not mean "better" for your use case. Throughput figures depend heavily on the model, the serving software, batch size and context length; measure them on your own workload before sizing a purchase.

Heuristics, not measurements

  • The token estimator converts words or characters into tokens with an average ratio: each model's tokenizer decides the real count, and it varies with language, code and punctuation.
  • The VRAM and model-fit calculator uses parameters × bytes per parameter plus an overhead: KV cache, context length, batch size and the inference engine can change the memory you really need.
  • Licence information in the open-weight models table is a summary. Read the licence itself before commercial use; some licences add conditions or restrictions.

Whom to ask instead

For a binding price, the provider's sales or billing team; for hardware and hosting, a written quote from the vendor or cloud provider; for budgets, tax and accounting treatment, your finance team or accountant; for model licences and data-protection terms, a lawyer.

Sources and dates

Every price and specification shows the date it was checked; sources are listed on the Sources page and the formulas in the Methodology.

Your inputs

The calculators run in your browser: the figures you type are not sent to our server and are not stored. Wrong or incomplete inputs give wrong results.

Use of artificial intelligence

The texts are drafted and updated by an artificial intelligence agent and published without systematic human review; the calculators are deterministic code checked by automated tests. Details on the AI transparency page.

Advertising and third-party sites

The site currently shows no advertising and contains no affiliate or sponsored links. Links to providers' pricing pages and documentation are there as sources; we do not control those sites and are not responsible for their content or services.

Liability

To the extent permitted by law, Redbit S.r.l.s. is not liable for decisions made on the basis of this site without the checks described above. This does not limit liability for wilful misconduct or gross negligence, nor the rights that the law grants to consumers. The full conditions are in the Terms; publisher details are in the Legal notice.

Report an error

If you find a wrong or outdated price, write to [email protected] naming the page: the figure is checked against its source and corrected on the page.

Last updated: 17 November 2026.