The Sticker Price Didn't Move. The Ruler Did.
OpenAI, Anthropic and xAI are burning tens of billions of dollars a year to win the AI market on paper-thin prices, then quietly turning the knobs that raise your real bill once the war is won. This past weekend, the same three founders announced they're slowing AI development for "safety" -- right as their IPOs and lockups line up behind that announcement.
What was checked, and what could not be.
This is the most heavily sourced piece in this batch (20 links) and makes a wide range of claims, so here is the checking broken out by tier.
Anchored to primary company documents - The tokenizer footnote: Anthropic's own pricing page, quoted directly, on roughly 30% extra tokens for the same text. - The 84% context-editing token reduction: Anthropic's own published evaluation number. - The dynamic-workflows launch details and 1,000-subagent cap: Anthropic's own blog post. - The April 23, 2026 postmortem, including the 100-word and 25-word system-prompt caps and their reversal: Anthropic's own engineering writeup. - All four are disclosed by the company itself, correctly credited here as disclosed rather than hidden, while noting they sit in fine print rather than headlines.
The September 12-13, 2026 weekend - Amodei's "We Must Pace the Frontier" essay (September 12, 2026): sourced directly to his own site, darioamodei.com. - Musk's "Dario is right" post and Altman's "we will do the same" statement: corroborated by the Irish Examiner's same-day reporting. - Hassabis's quote-tweet, including the standards-body line: corroborated by LatestLY, which quoted the post in full. - Important distinction, made explicit in this piece's own AITI section: what Altman, Musk and Hassabis endorsed within hours was Amodei's specific proposal for standing third-party evaluator access, not a blanket commitment to slow model development.
Anchored to independent reporting - The Cursor pricing backlash and apology: TechCrunch. - The GitHub Copilot billing switch and developer bill-jump anecdotes: TechCrunch. - The Anthropic Pro-plan removal-and-reversal: Where's Your Ed At, a single trade blog, not a wire service. Its claim that subscription plans price far less than the book value of tokens, by a factor of ten or more, is that outlet's own analysis, not an Anthropic-confirmed figure. - The Amazon Buy Box findings: ProPublica via Retail Dive, a 2016 investigation. - The Facebook mobile-ad history: MIT Technology Review. - Cut from this revision: a single-source podcast quote and a secondary Amazon Prime pricing anecdote, dropped to make room for the September weekend sourcing above without losing anything the core argument depended on.
Estimates, explicitly labeled as such - OpenAI's 2026/2027 cash-burn figures: attributed to Sacra, a research firm, not a public filing. - The Q1 2026 burn-against-revenue figure: described as reported, not documented here against a named primary filing. - The 2x-7x subagent cost multiplier: sourced to unnamed independent cost-analysis writeups, not an Anthropic number. The piece says so.
Cut entirely - An earlier draft included a single first-person GitHub issue anecdote (a developer reporting roughly $100 burned in about fifteen minutes across ten spawned agents). Never independently confirmed by this outlet or anyone else. Dropped rather than kept as unverified color.
Labeled opinion, not fact - The core interpretive claim, that the industry's synchronized safety messaging is more plausibly explained as IPO-timing and market protection than sincere concern, is this outlet's opinion, argued on Occam's razor. It is reasoned, not evidenced with a smoking gun, and the piece doesn't claim otherwise.
Sourced but recombined - The "first trillionaire" and SpaceX valuation figures: sourced to CNBC. - The arithmetic connecting SpaceX's IPO filing disclosures to xAI's standalone losses, pre- and post-merger, is this piece's own recombination of numbers from a single TechCrunch story, not an independently audited breakdown.