How accurate are OpenAI cost calculators, six measured against two sources
Every figure captured 29 August 2026. The pages measured can change at any time, so treat each reading as a dated snapshot, not a verdict on the site today.
We measured the OpenAI prices displayed by six cost calculators that rank for this topic, against two independent public price sources read the same day. Two of the six displayed rates that differ from both sources on the same model, with the widest gap at five times the corroborated price on GPT-5.6 Luna. Three more carried no rows at all for the current 5.6 generation. One page, the most recently dated of the six, matched both sources on every comparable row.
The practical point is not that any one site is careless. It is that model generations now ship faster than static price tables get edited, and a bill estimated from a stale or uncorroborated rate can be off by a multiple before a single token is spent.
How we measured this
Our method reads two independent public price sources in the same run, the OpenRouter models API and the LiteLLM price registry, and treats a rate as corroborated only when both agree within 1 percent. On 29 August 2026 the two sources agreed exactly on GPT-5.6 Terra, GPT-5.6 Luna, GPT-5.5, GPT-5.4, GPT-5.4 Mini, GPT-4o and GPT-4o Mini, and those models form the comparable set. Six ranking calculator pages were then fetched the same day, most twice over, a rendered fetch cross checked against the raw HTML, and every OpenAI price they displayed was recorded with the raw text around it. A displayed rate counts as differing only when it moves away from both sources on a model in the comparable set. Rates are USD per 1M tokens throughout. The raw per-site readings are embedded in this page's source as JSON, so the comparison can be re-run from the same captures.
Site by site, measured on 29 August 2026
invertedstone.com
Date on the page, no date label on the page. Comparable rows checked, 7. Rows matching both sources, gpt-5.5, gpt-5.4, gpt-5.4-mini, gpt-4o, gpt-4o-mini.
| Model | Displayed | Both sources say | Gap in and out |
|---|---|---|---|
| gpt-5.6-terra | $2.5 / $15 | $2 / $12 | 1.25x / 1.25x |
| gpt-5.6-luna | $1 / $6 | $0.2 / $1.2 | 5x / 5x |
The two gaps repeat in the page prose, which describes GPT-5.6 Luna as a cost efficient model at $1/$6 per 1M tokens. Both public sources price Luna at $0.20 input and $1.20 output. A reader budgeting a Luna workload from this page starts five times too high.
spurnow.com
Date on the page, page badge reads Latest numbers as of October 2024, while its meta description says April 2025. Comparable rows checked, 2. Rows matching both sources, gpt-4o-mini.
| Model | Displayed | Both sources say | Gap in and out |
|---|---|---|---|
| gpt-4o | $5 / $15 | $2.5 / $10 | 2x / 1.5x |
The gpt-4o row reads $5.00 and $15.00, which was that model's launch pricing. Both sources have carried $2.50 and $10.00 for a long time. The page also promises around 75 percent savings on cached input, while the current OpenAI generation publishes a 90 percent cached read discount, and its FAQ describes caching through a session identifier, which is not how the OpenAI guide describes prompt caching.
costlayer.ai
Date on the page, says latest published OpenAI pricing as of 2026, no day-level date. Comparable rows checked, 2. Rows matching both sources, gpt-4o, gpt-4o-mini.
The newest model on the page is plain GPT-5. The 5.4, 5.5 and 5.6 families are absent, so the freshness sentence and the model list disagree by three generations.
cloudzero.com
Date on the page, post dated July 02, 2026. Comparable rows checked, 2. Rows matching both sources, gpt-5.4-mini, gpt-5.4.
The rates it does print are right. The gap is coverage, the newest row is GPT-5.4 and the page itself says its calculator cannot model context window surcharges above the standard limit.
finout.io
Date on the page, badge reads Pricing Data Last Updated 2026-07-05, with Auto-checks monthly. Comparable rows checked, 3. Rows matching both sources, gpt-5.5, gpt-5.4, gpt-5.4-mini.
The rates in its live catalog are right. The catalog simply stops before the 5.6 generation, which shipped days after the badge's July check and had been out for weeks by our August capture.
costgoat.com
Date on the page, LAST UPDATED AUGUST 15, 2026. Comparable rows checked, 5. Rows matching both sources, gpt-5.6-terra, gpt-5.5, gpt-5.4, gpt-5.6-luna, gpt-5.4-mini.
Credit where due. Every comparable row matched both sources exactly at our capture time, and the page carries a plain dated label. Its GPT-5.6 Sol row is the one number we could not corroborate, and that is not its fault, as the next section shows.
The one price nobody agrees on
GPT-5.6 Sol is a special case, and it is why a corroboration rule matters more than any single table. At our capture time, two of the measured pages displayed Sol at $5 and $30. The OpenRouter feed carried $2 and $10, and the LiteLLM registry carried $4 and $20. Three different published price pairs for one flagship model, at the same moment. Our own OpenAI cost calculator withholds Sol for exactly this reason, and the split readings are documented on the price source disagreement page.
Why displayed prices drift
Three patterns show up in the readings. Tables built once and not dated drift as models reprice, which is how a launch price like gpt-4o at $5 and $15 survives long after the live price fell to half on input and two thirds on output. Catalogs updated on a schedule stop before the newest generation, which is how three pages with correct rates still had no 5.6 row weeks after those models shipped. And promotional windows can put two honest sources on different numbers for the same model at the same time, which is the Sol case above.
The defensive habit for a reader is short. Find the date on the page before trusting a rate, and prefer a page that names where its numbers come from. A page with no date and no source can be right, but you cannot tell from the page.