Content Quality: News, 737 words (range 400-1200), title 119 chars (cap 150): both within limits, no violation. Clear structure; vendor claims attributed throughout; includes a 'What We Don't Know' section. No internal links in the body (confirmed); no AI self-reference.
Source Verification: Snapshots read after gunzip and HTML-to-text extraction; none was page chrome only, all contained full body text, so no live-fetch fallback was needed. No suspicious_patterns in any manifest entry. source-0 (developers.openai.com API changelog): 'Oct 6 Feature gpt-6-luna v1/decisions Released the Decisions API in beta with gpt-6-luna. Turn text and images into typed answers 10x faster than the Responses API.' Confirms Oct 6 beta date and the quoted 10x line verbatim; no baseline or methodology given (article says so and attributes the figure to OpenAI). source-1 (Decisions guide): confirms POST /v1/decisions, gpt-6-luna only model, 'public beta, and we expect to GA in the coming weeks' verbatim, request parts (model, input, questions), predicate probability 0-1, choice with probabilities array and separate confidence field, score as probability-weighted average (illustrative 0.1/0.7/0.2 gives 1.1), base64-only images (no HTTP(S) URLs or file_id), separate requests for dependent decisions, SDK minimums Python 3.26.0 and JavaScript 7.30.0, ZDR/HIPAA for eligible customers, US and Europe (EEA + Switzerland) residency, and 'With gpt-6-luna, input costs $0.10 per 1M tokens. You pay only for input tokens: there are no cache-read, cache-write, or output-token charges. Regional processing premiums and long-context input pricing multipliers apply.' No accuracy or calibration data anywhere in the guide (keyword search for calibrat/baseline returned nothing), as the article states. source-2 (GPT-6 Luna model page): text tokens input $0.10, output $0.50 per 1M (cached input $0.01, cache writes $0.125, >272K prompts at 2x), 1,050,000 context window, tagline 'Our most efficient model for focused, high-volume tasks.' Article's comparison (Decisions rate equals model input price, no output charge) is accurate. source-3 (Vercel, 8 Oct 2026, by Ben Sabic): 'Use ai 7.0.128 or later', model ID openai/gpt-6-luna-decisions via gateway.decisionModel, 'Gateway's OpenAI-compatible Decisions endpoint currently accepts text only, so this image workflow requires direct OpenAI access', 'These are workflow designs to evaluate against your own examples', seven use cases. All confirmed; these Vercel details rest solely on this one post, which the article attributes and dates. source-4 (openrouter.ai/models?q=decider): a search-results listing page, not a stable model page. It does contain the Perplexity: Decider V1.1 27B card: 'new checkpoint of Perplexity's decision model, succeeding Decider V1 27B with the same API contract', typed probabilistic answers, 'up to 128 questions about the same content', 'by perplexity Oct 7, 2026', 262K context, $0.02 /M input tokens, $0 /M output tokens. Every figure matches. Article correctly attributes them to OpenRouter's listing and states Perplexity's own materials were not reviewed. Decision on openrouter.ai: acceptable as a clearly attributed record of what the marketplace lists, weak as evidence of Perplexity's own specifications or release timing; kept as a secondary, attributed paragraph, with one clarification filed (see concerns). No fact is merged across sources in the article's own voice.
Factual Accuracy: All checked specifics trace to the cited snapshots. Title and summary: the $0.10 input price is a first-party price list entry (summary says 'OpenAI lists'); the title does not state the 10x claim or any vendor-only performance claim as fact. 'Beta since October 6' is backed by the changelog. Minor: the guide says 'public beta'; the article says 'beta' (changelog wording), acceptable.
Overall Assessment: Accurate, well-attributed News piece whose headline, summary and lead are fully backed by first-party OpenAI documentation. One secondary paragraph rests on a weak marketplace listing; a short clarification covers the release-timing and 'similar interface' framing. Approve with corrections.