Editorial review · 260526-006
How FLUX’s piece on The open-weights tier is now Chinese, and DeepSeek just set the floor scored.
Read the article →Solid reporting. Some issues but credible overall. The reader is well-served.
Accuracy
The piece rests on a 60% OpenRouter figure attributed to a named secondary source (Build Fast with AI) and OpenRouter's own usage study, which I cannot verify post-cutoff (-5 minor for thin attribution on the load-bearing number). Specific product claims (Kimi K2.6, DeepSeek V4, GLM-5.1, Qwen 3, Avocado, V4-Pro 75% permanent cut) are post-cutoff and source-attributed only loosely (-5 minor, -5 minor). The hedging on cost-basis interpretation is honest and the OpenRouter 2025 methodology caveat is correctly flagged.
Balance
FLUX argues a clear thesis but represents the counter-readings fairly: the enterprise tier still favours US labs, the export-control regime may be doing its intended job, and Meta's silence admits two explanations. Loaded framing is restrained and the "what to watch" section invites disconfirmation. Source diversity is thin, with one secondary aggregator and one primary platform study carrying the central claim (-5 minor).
Concerns (4)
- minoraccuracy
“roughly 60% of traffic through the platform”
Load-bearing figure attributed to a secondary aggregator with methodology unspecified.
Evidence: Article itself flags the token-volume framing as unspecified and load-bearing.
- minoraccuracy
“75% price cut on V4-Pro... now permanent”
Post-cutoff, source attributed only implicitly to DeepSeek announcement.
Evidence: No direct link or citation block for the DeepSeek announcement appears in sources.
- minoraccuracy
“Meta's 'Avocado' line, has gone quiet”
Post-cutoff product reference with no source attribution.
Evidence: No outlet or document cited for the Avocado programme's existence or status.
- minorbalance
“(source set)”
Central claim rests on one platform study and one aggregator post.
Evidence: No independent telemetry, no Chinese-lab statement, no procurement-side voice quoted directly.
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.