AI//COST is an independent price desk that dates each LLM API rate from the vendor's own price page, for people comparing what a model will bill.
The homepage rate is a blend, and the date is on the card
The homepage flagship strip, read on October 28, 2026, says its rates were verified on 2026-10-23. Claude Opus 5.5 is $4 per million input tokens and $20 output, blended to $2.58, with a 1 million token context and cache reads at $0.20. GPT-6 Astra is $10 and $50, blended to $6.82. Gemini 4 Argon is $2 and $10, blended to $1.21. Grok 4.7 is $2 and $6, blended to $1.19, on a 500 thousand token context. Kimi K3 is $3 and $15, blended to $1.92. GLM-5.3 is marked cheapest at $1.40 and $4.40, blended to $0.78. The page says the blend assumes 92% input, 8% output, and 82% of input served from cache, and that this same assumption sits behind every effective-rate figure on the site. Where a vendor publishes no cached-input rate, no cache discount is applied, so that blend is a ceiling. The image on this page is that strip. These are the site's figures, not a price check run from here.
The Opus 5.5 page carries a different stamp: last verified October 19, 2026, against platform.claude.com. It says Opus 5.5 replaced Opus 5 as Anthropic's recommended default on September 22, 2026, at $4 and $20 against Opus 5's $5 and $25. Cache reads on this model are 5% of input, while the page says every other Opus reads cache at 10%. Thinking cannot be switched off, so reasoning tokens bill as output.
Three coverage counts, on three pages
The about page says the site tracks 340 LLM models across 35 providers, each number checked by hand against the vendor's own pricing page, with no scrapers and no third-party aggregators. A missing number stays blank. The calculator page says its tools run on 158 models across 33 providers. The homepage's own "how we work" block still says pricing is verified for 23 frontier providers, that the first calculator is live, and that side-by-side comparisons are next. That same homepage already shows comparison cards, including ChatGPT Free against Paid and Grok against Claude.
Calculators, and a public corrections log
- The calculator hub says there is no signup and no paywall. The tools cover monthly API spend, a token counter, subscription plans against pay-as-you-go, video clips, text-to-speech, embeddings, fine-tuning, and what one agent task costs once the context grows on every step.
- The corrections log records material fixes with the date, the page, what was wrong, and what is right. On October 13, 2026 it says the Nova 2 Lite page had priced single-region Geo inference at about $0.525 and $4.375 per million, and had said Amazon published no cache-read price. The log says the AWS Price List puts Geo (US) 10% above global, at $0.33 and $2.75, that $0.525 and $4.375 is the Priority tier, and that cache reads are $0.075. It says the global input and output prices had been correct, and that an internal audit flagged it.
- The methodology is labeled v1.3, with its own last-verified line on July 20, 2026, and says the rules are reviewed quarterly. It says a reader can write with a URL, and that they answer within 5 business days. Minor typos are not logged.
Who signs the prices
The author page names Yaroslav Vikhariev as founder. It says he started AI Cost & Tools Hub in 2026, and that every price is checked against the vendor's own pricing page on the date stamped at the bottom of that page. The page links his LinkedIn. The footer says some links are tracked and sponsored slots are marked. The about page says coverage is not conditioned on an affiliate commission. A weekly email digest sits at the bottom of the homepage.