Technology & Science
Anthropic Launches Claude Opus 4.7 With Enhanced Coding but Curbed Cyber Powers
On 16 Apr 2026 Anthropic began shipping Claude Opus 4.7 to the public, replacing 4.6 across partners and GitHub Copilot with double-digit coding and vision gains but a newly embedded filter that throttles offensive cybersecurity uses.
Focusing Facts
- On SWE-bench Pro, Opus 4.7 scored 64.3 %—up 10.9 points from 4.6 and ahead of GPT-5.4’s 57.7 %.
- Pricing holds at $5 per million input tokens and $25 per million output tokens, while an updated tokenizer now inflates token counts by roughly 1.0–1.35×.
- Cyber functions are gated: only vetted professionals may access them through Anthropic’s new Cyber Verification Program, while the public model auto-blocks high-risk security queries.
Context
Gate-k ept releases like Opus 4.7 echo the 1946 U.S. McMahon Act, which shared nuclear know-how only with trusted allies; frontier AI capability is likewise becoming a state-or-corporate-controlled resource. Over the past decade models have leapt from GPT-2’s 1.5 B parameters (2019) to Mythos-class systems that, per leaks, penetrate 32-step network defenses. Opus 4.7 marks a new phase: not raw performance, but selective downgrading—Anthropic deliberately shaved cyber skills while boosting coding and vision—to manage systemic risks. This suggests a future where top-tier cognition is withheld, spawning a stratified AI landscape reminiscent of mainframe access in the 1960s before personal computing democratized processing power. Whether this centralized stewardship averts catastrophe or merely entrenches power will shape the next century of technological governance.
Perspectives
Mainstream tech and business outlets
e.g., Mashable, Inc., GitHub Blog — Portray Claude Opus 4.7 as a major public-facing leap that beats GPT-5.4 and Gemini 3.1 Pro on coding benchmarks and expands multimodal abilities while remaining safe for general release. Their upbeat coverage depends on continued access to Anthropic and readership interest in the ‘next big thing,’ so they gloss over higher token costs and the fact that Mythos remains superior but locked away.
Skeptical tech press and commentator blogs
e.g., Decrypt, CNET, Techmeme — Argue that Opus 4.7 is merely a stop-gap; its real selling point is withheld Mythos, and the new model guzzles tokens and still trails GPT-5-4-Pro with tools. They lean into contrarian takes and FOMO to drive clicks, highlighting worst-case costs and corporate gate-keeping even when benchmarks show genuine gains.
International and regional news outlets focused on cybersecurity risks
e.g., GEO TV, Digit — Stress that Anthropic intentionally crippled Opus 4.7’s hacking skills as an experiment in safety after Project Glasswing, keeping Mythos behind closed doors for fear of cyber misuse. By centering the security angle, they may overstate the danger and underplay the model’s everyday utility to fit narratives about AI threats and regulatory urgency.
Like what you're reading?