Alberto Romero, writing at The Algorithmic Bridge, examines Anthropic’s April 2026 Opus 4.7 release, noting that the company simultaneously published benchmark data for a second model — Mythos Preview — which it said it would not release publicly due to cybersecurity risks.
Note: The central thesis of this article, described by Romero as “the non-democratic era of AI,” is behind a paywall and was not available to this publication. This article reports only on the verifiable factual claims visible in the public portion of the piece.
What it says
Romero describes Opus 4.7 as a 0.1 increment over Opus 4.6, reporting that Anthropic’s published benchmarks show the new model outperforming Opus 4.6 on 12 of 14 metrics Anthropic shared. He characterizes the model as oriented toward agentic and coding tasks rather than creative writing.
Romero notes that Anthropic released a comparison chart showing Mythos Preview outperforming Opus 4.7 on 11 of 11 benchmarks, while stating the model would not be released publicly. He describes the company as providing access to Mythos only to a limited set of industry and government partners, citing cybersecurity concerns. He writes that he had previously covered Mythos and found it “weird that Anthropic made it public at all, only to hold it back from us.”
Romero also comments on the pace of Anthropic’s releases, noting that Opus 4.7 arrived approximately two and a half months after Opus 4.6. He frames the accelerating release cadence as partly attributable to recursive self-improvement — models playing an increasing role in training subsequent generations — though he attributes this claim to the frontier labs generally rather than to Anthropic’s stated position.
Context
Romero states he was hearing early reports that the model “is worse than expected” but notes that firsthand experience is the only valid basis for assessing new models. The substantive argument of the piece — about what the Mythos decision implies for consumer AI users — is paywalled and was not available for this report.