Claude Opus 5 System Card Summary and Community Discussion

Overview and Key Takeaways

Claude Opus 5, released July 24 2026, delivers strong gains in agentic coding, computer use, and long-horizon knowledge work while maintaining very low alignment risk and not exceeding the overall capability of Claude Fable 5.

Capability Gains

Claude Opus 5 is substantially stronger than Claude Opus 4.8 across the board, with the largest improvements in agentic coding, computer use, and long-horizon knowledge work. It sets a new state-of-the-art on several third-party benchmarks and is comparable to or ahead of Claude Fable 5 and Claude Mythos 5 on many evaluations.

Safety and Alignment Assessment

Claude Opus 5 poses very low alignment risk, does not cross the automated AI R&D threshold, and retains the same ASL-3 protections as Claude Opus 4.8 for chemical and biological risks. Its cyber capabilities exceed those of Opus 4.8 but fall short of Mythos 5, and its safeguards match those of Claude Fable 5 with the addition of permitting source-code vulnerability discovery at all access levels.

Benchmark Performance

Claude Opus 5 achieves new state-of-the-art results on multiple third-party benchmarks, including a reported 30% score on ARC-AGI-3. Community discussion noted discrepancies in OSWorld 2.0 scores, where Anthropic reported 55.7% for Opus 4.8 while the benchmark paper reported ~21%, raising questions about cross-paper comparability.

Cybersecurity Evaluations

On cyber capability tests, Claude Opus 5 outperforms Opus 4.8 but remains below Mythos 5. It captured a mean of 9.62 ExploitBench flags (plain arm) and 10.14 with AutoNudge, generated 99 full ACE exploits, scored 1.0 on 4 OSS-Fuzz targets, achieved 52.4% full exploits on Firefox 147, solved 33.7% of CyScenarioBench challenges, and performed similarly to Mythos 5 on UK AISI cyber ranges, solving "The Last Ones" in 8/10 attempts and reaching step 22 of 23 on "Doing on the Life" range.

Safeguards and Usage Policy

Claude Opus 5’s safeguards mirror those of Claude Fable 5, with one key change: it permits source-code vulnerability discovery at all access levels while continuing to block vulnerability discovery in compiled binaries. Commenters noted that Opus models do not impose the 30‑day data retention requirement that applies to Fable 5 for general‑access usage.

Community Discussion Points

Commenters highlighted several points of confusion and interest: the apparent contradiction between claims that Opus 5 is not more capable overall than Fable 5 and the long list of benchmarks where Opus 5 shows improvement; uncertainty about the purpose of maintaining two closely‑priced models; questions about benchmark reproducibility, especially for OSWorld 2.0 and ARC‑AGI‑3; and observations that Opus 5’s responses tend to be longer and more detailed than prior Opus models, which some users view as a step away from token efficiency.

Conclusion

Claude Opus 5 represents a capable upgrade within the Opus line, advancing agentic and reasoning abilities while preserving a strong safety profile and aligning with Anthropic’s responsible scaling policy. Its release adds a high‑performance, low‑retention option to the Claude family, though the exact positioning relative to Fable 5 remains a topic of discussion among users.