Content Quality: Well-structured News piece (906 words by raw and link-stripped count; News range 400-1200). Title is 148 characters, counted with wc -m and Python; under the 150 cap, no violation. Leads with the facts, reports four inter-page discrepancies between Meta pages (preview vs "available today"; "1M tokens" vs 1,048,576; PDF missing from the model card modality list; "Muse Code subscribers" vs the Power Usage and Maximum Usage plans), has a candid What We Do Not Know section, and reproduces no benchmark figures (score tables are images). Only weakness is the closing Analysis paragraph (see concerns).
Source Verification: All six manifest entries have a non-null file with status 200; all read from disk (gunzip + text extraction). source-0 (about.fb.com newsroom, 9.6k chars of text) confirmed: "our most capable models to date", "state-of-the-art performance on personal and professional agentic tasks", Muse Wonder available today through the Meta Model API and Muse Code, Wonder Full available today for Muse Code subscribers, Wonder to power Muse soon and to subscribers first, full safety report "will be available soon". source-1 (research.meta.ai blog, fully server-rendered, 9k chars): same availability statement; score tables are graphics, so the article correctly reproduces none; Muse Spark 1.3 vs Wonder Full out-of-distribution plot is not used. source-2 (model card): the snapshot is a raw PDF stored in the .html.gz container (content_type application/pdf), not HTML; I extracted it with pdftotext (23k chars, full body, not a shell) and confirmed: release date October 12, 2026; modalities Input text, image, video, audio / Output text (no PDF); Context Length "1M tokens at inference"; results at max reasoning effort vs GPT-6 Astra, Claude Fable 5.1, Claude Opus 5.5; "the highest of our internally validated run, self-reported score, or official leaderboard score"; competitor runs "remain best-effort"; SWE-Marathon "this row is not a controlled common-harness comparison"; MemoryBench "an internal benchmark"; risk assessment is for Muse Wonder Full; without mitigations Meta does "not rule out" the high risk threshold in Chemical & Biological; Cybersecurity and Loss of Control "moderate or lower" unmitigated, "moderate or lower" across all domains with mitigations; ExploitGym 92 flags, 21 intended vulnerability, 71 unintended; limitation "Visual and spatial reasoning, especially 3D tasks, remains a known limitation." All verbatim. source-3 (dev.meta.ai/docs/models): "Muse Wonder is in preview"; muse-wonder-1.0 "Available (Standard mode only)" on Model API; Muse Wonder Full "Not available" on the Model API, "available only in Muse Code with a subscription", Power Usage and Maximum Usage plans, Muse Wonder on every plan; 1,048,576-token context; input text, image, video, audio and PDF; reasoning "minimal" to "max"; "long-running, complex tasks where capability matters more than cost per token"; Wonder and Spark "take the same requests on the same endpoints, so moving a workload between them is a model ID change". All confirmed. source-4 (dev.meta.ai/docs/pricing-rate-limits): Muse Wonder muse-wonder-1.0 Standard mode: cached input $0.20, input $4.00, output $20.00 per 1M tokens; Contributor mode "Not available"; "No long-context premium"; web search grounding $2.50 per 1,000 search queries; 3,000 RPM and 4,000,000 TPM; Muse Spark Standard $1.25 input / $4.25 output (cached $0.15). Title price claim is exact: it is the Standard-mode (only) rate for the Model API, not a preview-only or tiered price; no batch tier or long-context multiplier exists on the page, and the only omitted charge is web search grounding, which the article discloses. The page is undated; the article says it was read October 13. Arithmetic recomputed: 4.00/1.25 = 3.2 and 20.00/4.25 = 4.71, so "about 3.2 times" and "about 4.7 times" are correct and the figures $4/$1.25 and $20/$4.25 are the right ones. source-5 (TechCrunch, Russell Brandom, Oct 12): "the company's first attempt at a frontier model" (attributed to TechCrunch in the article; the TechCrunch headline phrase "first frontier model" is not asserted in the article voice), "comparable benchmark scores to Fable 5.1 and GPT-6 Astra" (per Meta's testing in the model card), "roughly equivalent in cost to GPT-6 Sol or Opus 5.5" (article attributes to TechCrunch and states rival prices were not checked), MemoryBench "higher than Fable, Opus, or Astra", "will only be available to Muse Code subscribers and will not be served through the API". All verbatim and correctly attributed. Title check: "Muse Wonder Full Is Limited to Muse Code Subscribers" matches the newsroom post and TechCrunch; the docs narrow it to two plans and the article reports that narrowing in the body, so the title does not overstate. Linked internal article /article/2026-08/06-meta-launches-muse-code-a-terminal-ai-coding-agent-built-to-take-on-claude-code-and-codex exists on main; it says Muse Code is powered by Muse Spark 1.2, beta in August, with $1.25/$4.25 pay-as-you-go rates; the description is accurate. Live re-fetch on 2026-10-13: dev.meta.ai/docs/pricing-rate-limits, dev.meta.ai/docs/models and the about.fb.com post all returned 200 and their extracted text is identical to the snapshots (no change since capture). No snapshot was missing or page-chrome-only, so no live-fetch fallback was needed.
Factual Accuracy: Every claim in Overview, What We Know and What We Do Not Know traces to a snapshot above. Vendor claims (state-of-the-art, benchmark standing, safety assessment) are attributed to Meta or TechCrunch; rival-model prices are attributed to TechCrunch and flagged as unchecked. One unsupported inference in Analysis: "the cheapest route to Meta's strongest model is a Muse Code subscription, not the API". Subscription prices appear in no cited page (the article itself lists them under What We Do Not Know), and Wonder Full is not offered on the API at all, so there is no cost comparison to draw. The preceding clause "The documentation draws a line that the announcement does not" is also loose, since the newsroom post and TechCrunch do state that Wonder Full is for Muse Code subscribers. Handled as a clarification.
Overall Assessment: Accurate, well-sourced and candid about discrepancies between Meta pages. Headline price and availability claims verified exactly; arithmetic correct; attribution of TechCrunch-only claims correct; no unsupported benchmark figures. One closing-paragraph inference is not supported by the sources and is covered by a single clarification. APPROVE_WITH_CORRECTIONS.