How We Source, Verify and Model LLM Pricing
Every number on this site comes from a provider's own pricing documentation, carries a verified date, and is re-checked weekly. This page explains exactly where the data comes from, how the calculators work, and — just as important — what our estimates leave out.
Primary sources only
We track eight models across three providers. Each rate in our data file carries two fields: the date it was last verified and the official page it came from.
| Provider | Models tracked | Official source | Last verified |
|---|---|---|---|
OpenAI GPT-5.6 Sol / Terra / Luna | 3 | OpenAI model documentation | 2026-08-23 |
Anthropic Claude Opus 5 / Sonnet 5 / Haiku 4.5 | 3 | Anthropic official pricing | 2026-08-09 – 2026-08-16 |
Google Gemini 3.6 Flash / 3.5 Flash-Lite | 2 | Gemini API pricing | 2026-08-16 |
We do not scrape resellers, aggregators or third-party pricing tables for our published rates. Where a provider's own documentation is inaccessible, we say so rather than substituting a secondary source. The provider invoice is always the source of truth.
A weekly cadence, with the date printed on every page.
Pricing in this market changes every few weeks — promotional windows open and close, rates get cut, scheduled increases get cancelled. A static price table is wrong within a month, so we re-verify every rate against the official documentation on a weekly cycle.
Each model carries its own verified date rather than one site-wide stamp, because we only update a date when that specific rate has been re-checked. Every pricing page prints the relevant dates in its header and in the calculator footer. When a rate actually changes, we update the data file and record the change with its effective date on the changelog.
Promotional rates are labelled as promotional wherever they appear, together with the date they expire and the rate that applies afterwards. If a page shows a promo price without saying when it ends, that is a bug — tell us.
How the token counter works — and why it is an estimate.
The on-site counters do not run each provider's tokenizer; that would mean shipping three large vocabularies to your browser. Instead they use a transparent weighted heuristic over character classes. You can reproduce every number it produces:
CJK characters (Chinese, Japanese, Korean) count 1.05 tokens each; emoji count 1.8; remaining letters and digits are divided by 4.05; punctuation counts 0.42; whitespace counts 0.08. The result is rounded and floored at one token.
That is accurate enough for budgeting and for comparing prompts against each other, and it is deliberately labelled an estimate everywhere it appears. It can differ from a real tokenizer by several percent, and the three providers disagree with each other too — the same document meters differently on OpenAI, Anthropic and Google. For anything billing-facing, use the provider's own tokenizer or usage API.
The two pricing rules we do apply
Our calculators are deliberately conservative: they apply only the pricing rules that are documented clearly enough to model correctly, and they say so on the page when a rule fires.
| Rule | Applies to | What we do | Where it matters most |
|---|---|---|---|
Long-context surcharge Documented on GPT-5.6 | OpenAI GPT-5.6 Sol / Terra / Luna, prompts over 272,000 input tokens | Input rate ×2, output rate ×1.5, flagged as "long-context rate" in the results table | OpenAI vs Gemini, where Gemini applies no surcharge at all |
Separate input / output rates All tracked models | Every provider | Input and output are always billed separately and never averaged | API Cost Calculator — output is typically 83–89% of a balanced bill |
Anthropic and Google apply no long-context surcharge in their current published rates, so none is modelled. Context windows and max-output caps are enforced in the Context Window Checker, which reports fit rather than silently truncating.
Six things our estimates deliberately leave out.
Published per-million-token rates are the ceiling of what you will pay, not the floor. We exclude everything we cannot model reliably from public documentation, which means real invoices are usually lower than the numbers on this site.
Prompt caching — repeated prefixes bill at a steep discount on most providers, and is typically the single largest saving available. Batch APIs — commonly around 50% off in exchange for hours of latency. Tool and function-call tokens, image and audio input, regional premiums and negotiated enterprise tiers, and tax.
We also exclude one thing in the other direction: we do not attempt to model free tiers, trial credits or startup programmes. Every figure assumes standard on-demand paid-tier text usage.
Methodology questions
Where does GetMoreTokens get its pricing data?
Directly from each provider's own published pricing documentation — OpenAI model docs, Anthropic's official pricing page, and Google's Gemini API pricing page. Every rate in our data file carries a verified date and a source URL. We do not scrape resellers, and we do not copy second-hand pricing tables without checking the primary source.
How often is the pricing data re-checked?
Weekly. Every rate is re-verified against the official provider documentation on a weekly cycle, and the per-model verified date is printed on each pricing page. When a rate changes we update the data file and record the change on the changelog page.
Is the token count exact?
No. The on-site counters use a transparent heuristic — a weighted character-class estimate — not each provider's tokenizer. It is accurate enough for budgeting but can differ from a real tokenizer by several percent. Exact counts require each provider's own tokenizer or usage API.
What do your cost estimates exclude?
Prompt caching, batch API discounts, tool and function-call tokens, image and audio input, regional premiums, enterprise and negotiated tiers, and taxes. All published rates are standard on-demand text-token rates. Real invoices are frequently lower than our estimates once caching and batch discounts apply.
Does GetMoreTokens accept payment from model providers?
No. We are not paid by OpenAI, Anthropic or Google to list or rank models. Our comparisons are arithmetic — published rates applied to fixed token volumes — not quality rankings, and not sponsored placements.
If we are wrong, we want to know.
Pricing errors on a cost-comparison site are worse than useless, so we treat them as urgent. If a rate, context window, promo expiry or calculation on this site disagrees with a provider's current documentation, email hello@getmoretokens.com with the page URL and the official source, and we will verify and correct it on the next cycle — corrections are logged on the changelog so the record stays auditable.