Unconfirmed: posts on X say Anthropic has already trained three internal models after Claude Mythos Preview, but Automatica has not confirmed the underlying claim.
@pankajkumar_dev wrote on 6 November that METR's report shows an internal progression from "Claude Mythos Preview" to "Model 2 : trained in July 2026," "Model 3 : trained in Aug-Sep 2026," and "Model 4 : trained in Sep-Oct 2026" @pankajkumar_dev on X1. The same post also says: "And Model 4 is already scoring above 50% on ALL 15 AECI benchmarks it was tested on" @pankajkumar_dev on X1.
@vikktorrrre wrote later on 6 November that "Anthropic has already trained THREE newer models after Claude Mythos Preview," listing the same timeline: "Model 2: Trained in July 2026," "Model 3: Trained in Aug–Sep 2026," and "Model 4: Trained in Sep–Oct 2026" @vikktorrrre on X2. That post also says "Model 4 scored above 50% on ALL 15 benchmarks in Anthropic's internal capability index (AECI)" and adds that "METR is now concerned that the benchmarks aren't difficult enough to accurately measure how capable these newer models are" @vikktorrrre on X2.
A third post, from @xiax0603 on 6 November, makes a similar claim in Chinese, saying METR disclosed four internal models including "Model 2（2026年7月完成）," "Model 3（8-9月完成）," and "Model 4（9-10月完成训练）" @xiax0603 on X3. The same post says "Model 4在所用15个基准全部超50%" and also claims that "全报告对未来模型的预测段落被删改" @xiax0603 on X3.
What these posts show, and what they do not, matters here. The posts are public and meet Automatica's threshold for reporting on named X accounts, but they are not independent corroboration. Automatica can verify that the posts were made; it cannot verify from the available evidence that Anthropic trained the models described, that METR directly observed those results, or that the benchmark scores and dates are accurate.
What would confirm the rumour is a primary document from METR, Anthropic, or another directly involved party setting out the model names, training dates, benchmark definitions, and evaluation conditions. What would kill it is a direct denial, correction, or publication showing that the cited timeline or scores were misstated. In the material provided here, Anthropic and METR have said nothing directly.
There is limited same-claim background in Automatica's earlier coverage. On 5 November, Automatica reported that "a public X post claims METR obtained unusual access to Anthropic's internal model work and saw results from unreleased systems," while noting that "the available evidence provides no primary documents or independent corroboration" Automatica Press (earlier coverage)4. Later coverage on 6 November similarly said an AI Alignment Forum post claimed METR found faster Anthropic AI R&D progress, but that Automatica could verify only that the post existed, "not the underlying claims, data or methodology" Automatica Press (earlier coverage)5.
That leaves readers with a narrow conclusion: multiple X accounts are circulating the same alleged Anthropic model timeline and benchmark claim, all attributing it to METR, but the available evidence contains no primary report, no methodology, no model configuration details, and no company confirmation. For a performance claim about unreleased systems, those missing pieces are the difference between a rumor and evidence.