Fordel StudiosAnthropic's Claude Opus 4.7 makes a big leap in coding, while deliberately scaling back cyber capabilities
What Happened
Anthropic's new flagship model Claude Opus 4.7 delivers major improvements in coding tasks. During training, the company deliberately tried to reduce certain cybersecurity capabilities. The article Anthropic's Claude Opus 4.7 makes a big leap in coding, while deliberately scaling back cyber capabili
Our Take
Claude Opus 4.7 now solves 74% of SWE-bench tasks, up from 63% in 4.6. The model was explicitly trained to weaken its ability to generate exploit code, even as coding proficiency improved.
This matters for RAG pipelines using LLMs to auto-generate or patch code. Teams relying on GPT-4 for code generation are paying 3x more per token than with Haiku—yet seeing comparable accuracy here. Believing bigger models always code better is a costly mistake.
Teams shipping code-assist tools should switch to Opus 4.7 for production workloads instead of defaulting to GPT-4. Shops focused on security tooling can ignore this—reduced exploit generation is a feature, not a bug.
What To Do
Use Claude Opus 4.7 instead of GPT-4 for code generation in production because it's 3x cheaper and now matches performance on SWE-bench.
Builder's Brief
What Skeptics Say
Deliberately weakening a model's capability undermines trust in its security assessments. If Opus can't write exploits, it can't properly red-team them.
2 comments
they literally trained it to be WORSE at something on purpose. that's a first
so the coding gains are real but they kneecapped the security research use case. great
Cited By
React
Get the weekly AI digest
The stories that matter, with a builder's perspective. Every Thursday.