Detailed Analysis
Anthropic released Claude Opus 4.7 on April 16, 2026, as a generally available upgrade to its predecessor, Opus 4.6, generating public confusion about its relationship to the company's separately maintained and far more restricted Claude Mythos Preview model. CNET's coverage addresses this confusion directly, clarifying that despite circulating speculation — partly fueled by a source code leak that referenced both models alongside other tiers such as Sonnet 4.8 and Capybara — Opus 4.7 and Mythos are entirely distinct products occupying different positions in Anthropic's model hierarchy. Opus 4.7 is broadly accessible through Anthropic's API and platforms such as Amazon Bedrock, while Mythos remains gated to select partners, including Apple for its Project Glasswing initiative, owing to its advanced and higher-risk capabilities, particularly in cyber-related domains.
In terms of raw technical performance, Opus 4.7 represents a meaningful step forward in agentic and software engineering tasks. It achieves 64.3% on SWE-bench Pro, 87.6% on SWE-bench Verified, and 69.4% on Terminal-Bench 2.0, outperforming rivals including GPT-5.4 and Gemini 3.1 Pro on those specific benchmarks. The model also incorporates an updated tokenizer that generates between 1.0 and 1.35 times more tokens than previous versions, improving output depth on complex tasks. Anthropic has explicitly framed Opus 4.7 as optimized for production-ready, long-running agentic workflows that require less human supervision — a design philosophy oriented toward reliability and safety rather than maximum raw capability.
The distinction between Opus 4.7 and Mythos reflects a deliberate and increasingly visible bifurcation in Anthropic's product strategy. Anthropic itself acknowledges that Mythos is "broadly more capable" than Opus 4.7, yet restricts its availability precisely because of the safety implications that come with that expanded capability. This tiered access model — where the most powerful systems are shared only with vetted partners under controlled conditions — represents a calculated approach to capability deployment that Anthropic has been developing alongside its frontier research. The arrangement with Apple for Project Glasswing underscores that the company is willing to commercialize Mythos-level capability selectively, rather than forgoing it entirely.
The broader significance of this release lies in what it reveals about the competitive and regulatory landscape surrounding frontier AI. Anthropic is simultaneously racing to match benchmark performance against OpenAI and Google while managing a dual-track release strategy that distinguishes between "safe enough for general availability" and "too capable for unrestricted release." This approach mirrors, in some respects, how pharmaceutical or defense technologies are tiered by access level based on risk profile. As source code leaks increasingly expose model hierarchies that companies would prefer to manage quietly, public understanding of AI capability stratification is improving — forcing companies like Anthropic to communicate more explicitly about why a less-capable model is being released publicly while a more-capable one remains in restricted circulation.
Read original article →