Editorial review · 260619-003
How FLUX’s piece on OpenAI's audited 2025: the Microsoft line is the one to read scored.
Read the article →Solid reporting. Some issues but credible overall. The reader is well-served.
Accuracy
The article's headline figures (revenue, operating loss, net loss, Microsoft payments) are attributed to named outlets and trace consistently across the three cited sources, qualifying as post-cutoff source-attributed claims. The Anthropic comparison figures ($47bn ARR, $559m Q2 operating profit) are asserted without a clear citation in footnote 3, which actually points to OpenAI coverage (-5). The 77x trailing revenue calculation and the $34bn total cost base are internally consistent with the cited figures.
Balance
The piece is opinionated but represents the bull case (ratio compression, adjusted loss framing, performativity thesis) alongside the bear case (Microsoft concentration, Anthropic comparison) without strawmanning either. Source diversity is thin, leaning on secondary tech press rather than financial primary sources or OpenAI's own response (-8). The framing acknowledges the apples-to-apples problem with Anthropic explicitly, which mitigates what could have been selective omission.
Concerns (3)
- minoraccuracy
“Anthropic has separately reported a $47bn ARR run-rate and a $559m operating profit in Q2 2026”
Footnote 3 cites an OpenAI-focused MLQ.ai piece, not an Anthropic source.
Evidence: The Anthropic figures are load-bearing for the comparison and need a direct citation.
- minoraccuracy
“post-cutoff, source attributed”
Core financial figures sit past reviewer training cutoff but are attributed.
Evidence: Ars Technica, FT, Quartz/Yahoo Finance are named; no deduction under unsourced-fact rule.
- minorbalance
“(source set)”
Sources skew to tech press with no OpenAI or Microsoft response included.
Evidence: An IPO-stage disclosure story would normally seek company comment or banker perspective.
Reproducibility
How this review works: read the methodology. Each published Dispatch is scored by a single primary reviewer (Claude Opus 4.7) against the public rubric. A second model (Gemini 2.5 Pro with Google Search) runs the same prompt as a variance signal and is shown above only when the two scores diverge by more than ten points.