Provenance Record

Well-structured News piece (Overview, What Meta Says, The Benchmark Figures, Where the Sources Differ, Weights, What We Don't Know). Separates Meta's first-party figures from Artificial Analysis's independent ones, labels every vendor row, scopes pricing to the 200K tier, and reports disagreements without reconciling them. Body is 1,161 words by the pipeline's whitespace split (News limit 1,200); title is 148 characters (limit 150).

All 8 stored snapshots read from disk (sources/2026-11/<slug>/source-0..7.html.gz, HTTP 200, sha256 of decompressed bytes re-checked against manifest; source-1 is the evaluation-report PDF, extracted with pdftotext); source 8 (thenewstack.io) has file=null / 403. SCANNER: one suspicious_patterns entry, on source-3 (dev.meta.ai pricing-rate-limits), pattern system-prompt-reference, excerpt 'Meta injects a small amount of steering context into every prompt (a system prompt and related scaffolding)...'. Read in context (line in the rate-limits notes): a statement that injected steering tokens are not billed. Not an instruction and not addressed to a reader; FALSE POSITIVE, no claim rests on it. THE NEW STACK (403): live fetch with a browser User-Agent returned a Cloudflare 'Just a moment...' challenge (HTTP 403); Archive.org Wayback has no capture (available API returns empty snapshots). Read by the publisher's RSS feed https://thenewstack.io/feed/ (HTTP 200) on 2026-11-10 ~16:48 UTC, item for /meta-muse-ember-launch/ (author Frederic Lardinois, pubDate Mon, 09 Nov 2026 18:00:00 +0000, full content:encoded text); item sha256 b7ddca5aecb3e87bdb50c0212994979de4c0ab985a47d6239192974f13e76130, whole feed file sha256 0fd8ec555814b85c2f0ebfdce6b1c68e4a60640d4f2fd7983f639b86094547c0 (feed lastBuildDate 16:27:34 UTC). CONFIRMED the only claim attributed to it: 'Like with previous Muse model releases, Meta is keeping Ember's model weights private.' (article quote 'is keeping Ember's model weights private' is verbatim). The New Stack also reports text/images/audio/video input, a 1M context and 512K output, Muse Spark 1.3 at 37 and $2.39, and OSWorld binary scores; none contradicts the article. (0) research.meta.ai/blog/muse-ember (announcement, dated November 9, 2026): 'our smallest and most efficient model'; tool calls, coding subtasks, routing, extraction, long document parsing; 'matches Haiku's best score at about a third of the cost and beats every Luna setting'; chart data Muse Ember 33.9/$0.116 (xhigh) and 35.8/$0.136 (max), Haiku max 36.2/$0.411; caption 'Muse Ember costs are estimates at proposed pricing'; footnote 1 'cost per task uses each model's list pricing' and 'Intelligence Index v5' (NOT 'v5.0': see correction 2); AutomationBench-AA 37.8/20.2/13.9 marked 'Strict pass@1.'; Terminal-Bench 4.0 40.9/39.2/16.4 'best published harness' (mini-swe-agent, Claude Code, Codex); Coding Agent Index-AA 46/37/31; FrontierCode 34.9 Haiku vs 27.2 Ember vs 25.2 Luna; OSWorld partial 69.5/68.9/50.5 'Meta internal evaluation. Partial-credit reward.'; Agents' Last Exam 61.6/57.0/57.1 'Meta internal evaluation'; PSI 46.8/36.1/37.0 'Internal composite'; safety 'remained below the framework's high-risk thresholds' (curly apostrophe matches); 'first trained using Muse Wonder Full'; Wonder/Spark/Ember positioning; Spark 1.3 comparison 36 vs 37, $0.14 vs $2.39, $0.10/$0.50 vs $1.25/$4.25; footnote 2 over-200K rates $0.30/$0.012/$1.50 - all CONFIRMED. (1) evaluation report PDF (November 2026): AutomationBench-AA is 'Artificial Analysis's run of Zapier's private set of 657 business-workflow tasks', 'pass@1 (Artificial Analysis's Tasks Completed): the share of tasks completed in full, with every objective met and no guardrail violated'; Terminal-Bench 4.0 'Muse Ember's result is Artificial Analysis's run in the mini-swe-agent harness... For Claude Haiku 5.5 we use Anthropic's self-reported Claude Code result from the Claude Haiku 5.5 system card'; 'Internally generated competitor runs use common settings where practical but remain best-effort'; names 'Intelligence Index v5.0' - CONFIRMED. (2) dev.meta.ai/docs/muse-ember: muse-ember-1.0; input text, image, video, audio, PDF; output text; context 1,048,576; max output 524,288; reasoning 'minimal' to 'max', default max; Responses/Chat Completions/Messages; 'Muse Ember runs in Standard mode only; it has no Contributor-mode variant'; use Spark/Wonder for hard single-shot coding, knowledge-heavy, visual coding; available on Model API and in Muse Code - CONFIRMED. (3) dev.meta.ai/docs/pricing-rate-limits (snapshot taken 2026-11-10): Ember up to 200K: $0.004 cached / $0.10 input / $0.50 output; over 200K: $0.012 / $0.30 / $1.50; 'billed at the long-context rates in full, not only for the tokens past 200K' (verbatim); web search $2.50 per 1,000; muse-ember-1.0 3,000 RPM / 4,000,000 TPM; Spark Standard $1.25/$4.25; Wonder Standard $4.00/$20.00; Contributor mode definition; 'Contributor mode isn't available for Muse Wonder or Muse Ember' - CONFIRMED. (4) dev.meta.ai/docs/models: Ember among API-hosted lines (input text, image, video, audio, PDF); Muse Glimmer and Muse Spark Open open-weight; no Ember parameter count in any of the five Meta pages (the only parameter count on them is Muse Spark Open's 816B) - CONFIRMED. (5) Artificial Analysis article (Nov 9, 2026): 'the cheapest model scoring 36 or higher' at $0.14 per task (max); xhigh scores 34 for $0.12; Haiku (max) 'almost 3x more at $0.41 for the same score'; 'roughly a third as many output tokens'; cached input 60% cheaper; 'scores 68.1% on AutomationBench-AA, level with Muse Wonder (max)'; Terminal-Bench 4.0 40.9% vs Haiku 32.8% 'in our Intelligence Index evaluations', FrontierCode 27.2 vs 34.9; AA-Omniscience accuracy 21.1% vs 34.7%, declines 64%, hallucination rate 17.4% 'half of Claude Haiku 5.5's'; Coding Agent Index 46/37/31 - CONFIRMED. The article states no metric for 68.1% (searched for partial/strict/completed/pass@: no match). (6) Artificial Analysis model page: Index 36 'Muse Ember (Max)', $0.14 'Cost per Intelligence Index task' (weighted average), 'Proprietary model', 'supports text and image input', 448 tok/s, 3.1 minutes per task, Index v5.0.1 - CONFIRMED. A live curl with a browser User-Agent on 2026-11-10 returned byte-identical pages to both stored snapshots (sha256 prefixes 08de2b61d02b and d27bb3916bb0). The two AA pages AGREE with each other (36, $0.14, max, v5.0.1 on the model page; the article does not name an index version). Page source: Muse Ember (max) record holds automationBenchPartialScore 0.6807901586086508 (68.1%), Muse Wonder (max) 0.6809455131207758, GPT-6 Astra 0.6849174806688876; the article's '0.6808 under a field named automationBenchPartialScore' is accurate and labelled as the Herald's own reading of page source. (7) The Decoder (Matthias Bastian, Nov 9, 2026, Short News): $0.14 per task, Spark 1.3 $2.39 for 36 vs 37, FrontierCode 34.9 vs 27.2, text/images/audio/video, 1M tokens, $0.10/$0.50 - CONFIRMED. TITLE TEST. '36 at $0.14 per Task at Max Reasoning on Artificial Analysis': both AA pages and Meta (35.8 / $0.136, rounded by AA) agree; $0.14 is AA's weighted average per Intelligence Index task at list price. '68.1% on AutomationBench-AA Versus Meta's 37.8%': both figures are exact and correctly attributed (AA article; Meta table). Meta's report defines 37.8% as full-completion pass@1; AA does not name its metric and its page data calls the field a partial score. 'Versus' juxtaposes the two figures without saying they are on the same basis, but the headline asserts no equivalence and no cause, both numbers are labelled by whose they are, and the Overview and the first 'Where the Sources Differ' bullet disclose the metric question with a hedge. No wrong statement in headline, summary or lead, so not a REJECT; a reader of the headline alone could infer a like-for-like gap, so a clarification is filed (correction 1). CLAIM-ACCURACY TESTS. Quantifier/absence scan (both, every, all, only, none, neither, no source, first): 'Both sources give Muse Ember 40.9%' holds (Meta 40.9, AA 40.9); 'neither source's text explains the gap' holds against all nine sources (AutomationBench appears only in Meta's pages and AA's two pages; The Decoder and The New Stack do not mention it); 'None of the cited pages reconciles these' (input types) holds; 'beats every Luna setting' is Meta's claim, attributed and true against Meta's chart (Luna max 27.8 < 33.9); 'first model it trained using Muse Wonder Full' attributed to Meta. 'Differ on two benchmark rows': AutomationBench-AA and Terminal-Bench 4.0 (Haiku) are the only overlapping rows that disagree (GDP.pdf 22.8/20.8, FrontierCode 27.2/34.9, Coding Agent Index 46/37/31, Ember Terminal-Bench 40.9 all match). Dates: announcement November 9, 2026 is Meta's date and also AA's, The Decoder's and The New Stack's (pubDate 2026-11-09 18:00 UTC; Monday); no time-zone conflict. Derived figures: the article computes none; all vendor figures are labelled with their row, maker or version. Pricing is scoped to prompts up to 200K input tokens, with the over-200K tier and the read date November 10, 2026. Disagreements (Terminal-Bench Haiku, input types, cost basis, index version, AutomationBench) are each reported with attribution and left unreconciled, as policy rule 6 requires; checked that no source explains them. Licence/weights: only the three attributed statements (Meta models page, The New Stack quote, AA 'Proprietary model'). No unsourced quote found. HERALD CROSS-LINK: /article/2026-10/13-meta-releases-muse-wonder-on-its-model-api-... exists and resolves (published 2026-10-13); its corrections record (clarification on the 'cheapest route' sentence) concerns a claim the new article does not repeat; the new article cites only the Wonder $4.00/$20.00 rates from Meta's pricing page. Anthropic disclosure: not required - the subject is Meta's Muse Ember; Anthropic and Claude Haiku 5.5 / Claude Code appear only as a competing model in benchmark tables and as the harness provenance of a competitor's Terminal-Bench figure (not in the headline, summary or lead, not a finding or credit line), which the policy treats as an incidental name among competing models. ALLOWLIST: research.meta.ai, dev.meta.ai (lines 275-276), artificialanalysis.ai (2591), the-decoder.com (124) and thenewstack.io (119) are all in config/source_allowlist.txt; no additions. MarkTechPost and Simon Willison are not cited.

Every number, quote, date and version was checked against raw text and holds, with two clarifications: (1) the headline sets AA's 68.1% against Meta's 37.8% without saying the two are not shown to share a metric; (2) the article writes that Meta's footnote cites Intelligence Index 'v5.0' but the announcement's footnote reads 'v5' (Meta's evaluation report says 'v5.0'). The hedged page-source observation about automationBenchPartialScore is accurate (0.6808).

Accurate, well-sourced and carefully hedged News coverage of a model launch with real source disagreements; publishable with two clarifications recorded. Featured candidacy: 1 - solid, accurate coverage for a developer audience choosing small agent models; one launch in an existing model family, not industry-changing