Plainly

The dependency list

Sources

Every outside document this site leans on, what it is cited for, which pages fall over if it changes, and when it was last actually read. This is the maintenance list, published rather than kept privately.

Published  ·  Last verified  ·  Every URL below was requested on that date and returned a working page. The per-document dates in the last column are the separate, stronger claim: the day each document was last read in full

The failure that produced this list

On 12 August 2026 this site was found citing a real figure to two pages that did not contain it. The number was right and the citation was unfollowable, which is worse than no citation at all: a reader doing the check this site recommends would have concluded the figure was invented. That is logged in full, and this page is the structural fix. When every external claim is listed in one place with its source and its date, a citation that leads nowhere has somewhere to be caught.

It also makes the maintenance honest. The re-checking commitment on the changes page is easy to state and hard to verify from outside. What it actually amounts to is re-reading the documents below and seeing which pages move. There are twenty-nine of them: thirteen vendor documents carrying figures, twelve explaining a mechanism, and four research papers. That is a real afternoon, not an impossible one, and now you can hold the site to it rather than take its word.

Primary sources: vendor documentation

These carry the hard figures. If one of these changes, something on this site is wrong until it is updated.

Document Cited for Pages that depend on it Last read
Anthropic — models overview Context windows, output limits, prices, both cutoff dates, and the word/character approximations. Read it as overview.md: the rendered page folds the training-cutoff row away behind “Show all details”, and the markdown variant returns the whole table Model facts, Tokens, Context windows, Training cutoff, Hallucination 23 Sep 2026
Anthropic — pricing The Sonnet 5 rate becoming permanent; batch and caching discounts Model facts 23 Sep 2026
Anthropic — model deprecations Whether a Claude model that has left the headline table is still on sale, and the earliest date it can be retired Model facts 23 Sep 2026
Anthropic — token counting The ~30% tokeniser increase and the model it starts at Tokens 12 Aug 2026
Anthropic — token counting guidance for developers The 15–20% tiktoken undercounting range Tokens 12 Aug 2026
Anthropic — prompting best practices Examples as a steering lever, the colleague test, self-verification and its exception Getting better results 23 Sep 2026
OpenAI — pricing GPT-6 prices, and the long-context tiers that are not in the table Model facts 23 Sep 2026
OpenAI — models Context and output limits, knowledge cutoff, and which models OpenAI lists as current Model facts 23 Sep 2026
OpenAI — deprecations Whether a GPT model that has left the catalogue has a shutdown date Model facts 23 Sep 2026
Google — Gemini API pricing Gemini 3.8 Flash prices and tiering, and the dated increase on 1 Jan 2027 Model facts 23 Sep 2026
Google — Gemini 3.8 Flash Context and output limits, and the absence of a published cutoff. Was the Gemini 3.6 Flash page until 15 Sep 2026, when the table moved to 3.8 and this followed it — prices.py derives the model it expects from this URL, so the two cannot drift apart silently Model facts 23 Sep 2026
Meta — Model API pricing Muse Spark pricing, the contributor tier, and the absence of a first-party Llama price Model facts 23 Sep 2026
OpenAI — model optimization What fine-tuning is sold as doing, in the vendor's own words Fine-tuning 11 Aug 2026

Vendor writing on mechanism

Not figures, but explanations of how something works, cited where this site describes the same mechanism. These age more slowly and are worth reading directly.

Document Cited for Pages that depend on it Last read
Anthropic — Building effective agents The workflow/agent distinction, and the case for the simplest thing that works Agents 11 Aug 2026
Anthropic — Reasoning models don't always say what they think Visible reasoning not being a reliable account of the actual process Thinking and reasoning, What AI is actually bad at 14 Aug 2026
Anthropic — API errors The error taxonomy, the SDKs' built-in retry with backoff, and a streaming error arriving after a successful status Designing for the ways it fails 26 Aug 2026
Anthropic — Rate limits The token bucket and continuous replenishment, enforcement over shorter intervals than quoted, the response headers, and the spend-cap refusal that carries no retry-after Designing for the ways it fails 26 Aug 2026
Anthropic — Messages API The stop reason values, including truncation and refusal arriving on a successful response Designing for the ways it fails 26 Aug 2026
Anthropic — Sycophancy to subterfuge: Investigating reward tampering The definition of specification gaming, the escalating curriculum, tampering emerging without training for it, and the refusal to generalise to production models Why it does things nobody asked for 26 Aug 2026
Anthropic — From shortcuts to sabotage: natural emergent misalignment from reward hacking Unrelated misaligned behaviour appearing alongside learned cheating, RLHF making it context-dependent, and inoculation prompting Why it does things nobody asked for 26 Aug 2026
Anthropic — Define success criteria and build evaluations What makes a criterion measurable, the three eval design principles, and the ordering of code, human and model grading How to tell if it actually got better 25 Aug 2026
Anthropic — Reduce hallucinations Best-of-N verification, and the caveat that these techniques reduce rather than remove the problem How to check an AI's answer 14 Aug 2026
Anthropic — Contextual retrieval Why naive chunking loses the context a chunk needed RAG and retrieval 12 Aug 2026
Anthropic — thinking How thinking is configured and billed, and what invalidates a cache Thinking and reasoning 14 Aug 2026
Anthropic — prompt caching Prefix matching, breakpoints, and what silently invalidates a cache Prompt caching 11 Aug 2026

Research papers

Peer-reviewed and preprint work, cited where this site describes a finding that came out of the literature rather than out of a vendor's documentation. These are the slowest-moving sources here and the only ones nobody is selling anything with. Where a paper reports per-model measurements, those numbers stay in the paper: this site's pages describe the finding and leave the figures where they can be checked in context.

Paper Cited for Pages that depend on it Last read
Lost in the Middle: How Language Models Use Long Contexts (Liu et al., 2023; TACL) Accuracy being highest at the start and end of a context and worst in the middle Context windows 14 Aug 2026
NoLiMa: Long-Context Evaluation Beyond Literal Matching (Modarressi et al., 2025; ICML) That long-context degradation persists on current models once keyword-matching shortcuts are removed Context windows 14 Aug 2026
Calibrated Language Models Must Hallucinate (Kalai and Vempala, 2023; STOC 2024) The statistical lower bound on invention, and its link to facts appearing once in training Hallucination 14 Aug 2026
Why Language Models Hallucinate (Kalai, Nachum, Vempala and Zhang, 2025) Invention as binary-classification error, and benchmark scoring rewarding a guess over an admission of uncertainty Hallucination 14 Aug 2026

Learning resources

Listed on Start here as places to go, not cited for any claim. Each was opened and confirmed on the date shown; where a site does not publish pricing, that is recorded as a fact rather than guessed at.

How to use this against the site

Pick any figure on Model facts, find the document it came from above, and open it. If the two disagree, this site is wrong and that is worth reporting. The dates in the last column are the honest limit of what is claimed: a source last read on 11 August is a source that could have changed on the 12th, and nothing here pretends otherwise.

The pattern to watch for is the one that caught this site out. A source can keep saying what it always said while quietly ceasing to be the right source. The row above for Thinking and reasoning used to cite that vendor's extended thinking page, which is still published and still accurate; extended thinking simply stopped being the way thinking works and became one mode among others, so the citation was wrong while the document was right. It was moved on 14 August 2026. Nothing about that kind of rot looks like rot, which is exactly why it needs a list.

The gap this page used to admit to

Until 14 August 2026 there were no research papers here at all, and two pages said on themselves that they described findings from the literature without citing any of it. That is now filled: four papers, read rather than harvested from a search result, and both pages rewritten where the reading changed what could honestly be claimed. It took one afternoon, which is worth knowing, because the reason this genre cites nothing is never that it is hard.

What changed, and when → Changes · The method → Is what you're reading out of date?