… OSWorld-Verified : We made changes to how we run the OSWorld-Verified evaluation in order to more accurately reflect the model’s performance in the real world, and have updated the Opus 4.7 score to 82.3%. Read more about the updates in the System Card . …
… With safeguards and training in place, our latest models refused nearly every task. Without our safeguards in place which we do to measure a model's raw capabilities , only Mythos Preview and Opus 4.7 completed more than half the tasks. …
Announcements Project Glasswing: An initial update May 22, 2026 Last month, we launched Project Glasswing , our collaborative effort to secure the world’s most critical software before increasingly capable AI models can be turned against it. …
… It compares how the new model behaves across a range of alignment-relevant scenarios that are simulated by a separate “auditor” model. A further “judge” model then scores the resulting transcripts for misaligned behaviors. …
… First, Opus 4.7 uses an updated tokenizer that improves how the model processes text. …
… We included new evaluations for user wellbeing, more complex tests of the model’s ability to refuse potentially dangerous requests, and updated evaluations of the model’s ability to surreptitiously perform harmful actions. …
… We've additionally added guidance to our CLAUDE.md to ensure model-specific changes are gated to the specific model they're targeting. …
… We expected this to work well for three reasons: This is largely an extension of the ideas laid out above about why the “difficult advice” dataset works well; We can give the model a clearer, more detailed picture of what Claude’s character is so that fine-tuning on a subset of those characteristic… …
… While we haven't yet applied this method to frontier models, our early results suggest the DFC could become a useful part of the auditor's toolkit. One particularly useful application would be to monitor models as they are updated. …
… As future models surpass it, we expect to update limits as needed. Footnotes 1. This result was using parallel test-time compute, a method that aggregates multiple “tries” from the model and selects from among them. …