Unconfirmed: posts on X are circulating a claim that a model called “Obsidian 3” appeared in FrontierCode 2.0 results and scored above Anthropic’s Opus 5.5, but Automatica has not confirmed the model, the benchmark entry, or who operates it.
@CodexResets1 wrote on 28 October that “Anthropic’s next model may have just leaked” and claimed that “FrontierCode 2.0 features a mystery model, ‘Obsidian 3,’ scoring 46.8% and outperforming Opus 5.5’s 45.5%” @CodexResets1 on X1. In the same post, the account added: “The Claude Code harness makes Anthropic a strong suspect” and asked, “Could this be Fable 5.5 or Opus 5.6?” @CodexResets1 on X1.
Those are the core claims in the rumour, and the available evidence is thin. The available evidence contains no primary benchmark document for FrontierCode 2.0, no Anthropic announcement, and no independent testing. That means the numbers, model identity, and implied vendor attribution are all unverified from the material reviewed here.
A second X post in the available evidence does not mention Obsidian 3 directly, but it does show the same day’s broader dispute over model rankings. @superalesha wrote on 28 October that a new index allegedly placed “Claude Haiku 5.5” above “DeepSeek V4.1 Flash 1017 and GLM-5.3,” then argued that benchmark changes, including “private tests nobody outside can see,” made the ranking unreliable @superalesha on X2. That post is relevant only as background on why social benchmark claims are being challenged; it is not confirmation of the Obsidian 3 claim.
What would confirm the rumour is straightforward: a public FrontierCode 2.0 results page showing “Obsidian 3,” a methodology note identifying the evaluation conditions, or an Anthropic statement acknowledging the model or the benchmark run. What would weaken or kill it would be equally clear: a benchmark correction, evidence that the entry belongs to another lab, or documentation showing the score or harness attribution was misread.
So far, the companies involved have said nothing in the material provided here. The available evidence includes no statement from Anthropic and no primary FrontierCode operator documentation.
The closest relevant background in Automatica’s earlier coverage is a separate unconfirmed report on benchmark methodology. On 28 October, Automatica reported that posts were claiming an Artificial Analysis “Intelligence Index v5” update widened the apparent gap between open and closed models, while noting that Automatica had not confirmed the update, methodology, or quoted scores from primary documentation Automatica Press3. That does not support the Obsidian 3 leak, but it does show that benchmark interpretation is itself under dispute.
Given the evidence available, readers should treat this as a source-attribution story, not a performance finding. One X account says a mystery model beat Opus 5.5 on a named benchmark; nothing in the available evidence independently establishes that the model exists, that Anthropic made it, or that the score is directly comparable under documented test conditions.