Claude Opus 5 system card says coding and computer use jumped, while cyber limits remain
Don't Worry About the Vase (Zvi) · rss · 2026-07-25
Zvi’s write-up says Claude Opus 5 is substantially stronger than Opus 4.8, with the biggest gains in agentic coding, computer use, and long-horizon knowledge work.
Key points from the system card discussion:
- Opus 5 is described as faster and cheaper than Anthropic’s larger models, while still reaching or approaching top-tier benchmark results.
- Anthropic says it is weaker than the most capable “Mythos”/frontier model on the most dangerous cyber tasks, partly because it was deliberately not trained on cyber data.
- Safety filters were changed: default safeguards now resemble Fable 5, and vulnerability discovery in source code is allowed, while compiled binaries remain blocked.
- The write-up questions whether some of the reported safety and alignment improvements are actually robust, especially given saturated autonomy evals.
The post also cites Anthropic claims that classifiers now trigger far less often, reducing false positives and making the model less annoying for normal users.
More from Models
- Benchmark scores drop from 89% to 19% on new evals — how benchmaxxing breaks leaderboard trust — airesearch12 · 2026-09-11
- ChatGPT tells user their question is too hard and to 'accept dumber answers' — phido3000 · 2026-09-11
- Developer Building a Unified Leaderboard of All Model Benchmark Scores — airesearch12 · 2026-09-11
- Rumor claims Kimi faked performance by serving Claude; DeepSeek new model surprises in evals — realsohamparekh · 2026-09-11
- GPT-5.6 writes well but is instantly forgettable, user complains — BasedRaddka · 2026-09-11
- Opus Refuses Protein Research Codebase Over 'Safety' Concerns, Dev Considers Rolling His Own — josephdviviano · 2026-09-11