Technology & Science
Anthropic Releases Claude Opus 4.7, First Public AI Deliberately Downgraded for Cyber Safety
On 16 Apr 2026 Anthropic began offering Claude Opus 4.7 worldwide—an upgrade that outperforms GPT-5.4 in coding but whose offensive-security skills were intentionally reduced under Project Glasswing, making it the first large-scale “frontier” model gated behind identity checks and real-time cyber-request blocking.
Focusing Facts
- Opus 4.7 scored 64.3 % on the SWE-bench Pro coding benchmark, topping GPT-5.4’s 57.7 % (Anthropic data, 16 Apr 2026).
- The model now accepts images up to 2,576 px on the long edge (~3.75 MP), more than triple Opus 4.6’s limit.
- Pricing held steady at $5 / M input tokens and $25 / M output tokens, despite up to 35 % higher token usage.
Context
Tech firms have rarely admitted to crippling a flagship product for safety; the last mass-market parallel is the 1992-1999 U.S. export limit of 40-bit web encryption that created a two-tier internet until policy caught up. Anthropic’s self-imposed throttling echoes that era’s fear of unleashing tools that adversaries could weaponise faster than regulators can react. Long-term, this marks a pivot from “open by default” AI toward licence-based, dual-use governance that resembles how nuclear know-how (post-1945) and CRISPR gene-editing (after 2012) were cordoned off for vetted labs. Whether Opus 4.7’s gated release becomes the norm or an historical footnote will shape who controls general-purpose intelligence a century from now: a public utility or a permissioned technology cartel. Either way, the event signals that capability progress is now entwined with safety bureaucracies, foreshadowing an AI landscape regulated as tightly as aviation or pharmaceuticals by 2126.
Perspectives
Developer-centric tech blogs
e.g., iClarified, Windows Report, GitHub Blog — Portray Opus 4.7 as a standout technical upgrade that overtakes GPT-5.4 on coding benchmarks and offers developers sharper vision and finer control. Coverage is product-boosterish, focused on benchmarks and ecosystem tie-ins while glossing over the model’s intentionally reduced cyber powers because their readership cares most about practical capabilities.
Skeptical consumer tech press
e.g., Tom’s Guide, Gizmodo — Cast Opus 4.7 as a purposely "nerfed" public release that mainly serves to spotlight the secretive, more dangerous Mythos model and tighten access through Project Glasswing. Uses a cautionary, almost alarmist framing that highlights secrecy and risk to attract attention, risking under-valuing genuine advances the model still delivers.
Financial market outlets
e.g., Investing.com, Quartz — Read the launch through a business lens, stressing how Anthropic’s new model and design push could disrupt incumbents’ share prices and cement big-tech partnerships. Share-price and competitive angles dominate, which can exaggerate near-term market fallout and leave technical or safety details under-explored.
Like what you're reading?