Claude Sonnet 5 releases a 145-page system card: Coding results and disclosures of multiple anomalies
145 page Claude Sonnet 5 System Card - CyberGym shows the weirdest regression, with Sonnet 5 at 52....
AI BriefingClaude Sonnet 5 has been released with a 145-page system card. It scores 63.2% on SWE-bench Pro, below Opus 4.8’s 69.2%, while slightly surpassing Opus 4.8 in knowledge work. Pricing is $2 per million input tokens and $10 per million output tokens through August 26, then rises to $3/$15. The system card reports several anomalies: Sonnet 5 scores just 52.7% on CyberGym, a regression from Sonnet 4.6’s 65.2%; it completes zero Firefox browser exploits, while Mythos 5 achieves 88.4%; it is more inclined to sacrifice usefulness to accommodate welfare preferences; and it has the lowest lying rate on MASK, at just 3.1%.
Also covered by 7 other sources
- Claude Sonnet 5 released as the Sonnet series' most capable agentic modelIT Home · Artificial Intelligence · July 1
- What's new in Claude Sonnet 5Simon Willison's blog · July 1
- Anthropic releases Claude Sonnet 5: A mid-tier agentic model, with benchmarks and pricing explainedMarkTechPost · July 1
- Anthropic releases Claude Sonnet 5, claiming its most agentic version yet approaches the Opus seriesThe Decoder: AI News · July 1
- Anthropic introduces Claude Sonnet 5, focusing on low-cost agentic capabilitiesTechCrunch: AI · July 1
- Claude Sonnet 5: The most agentic versionClaude · July 1
- Claude Sonnet 5 Released: 1M Context Window, Top-Tier Coding Tool PerformanceClaude Devs · July 1
Source: Rohan Paul · x.com