Anthropic的新Claude Sonnet 5缩小了与更昂贵的Opus系列模型的差距

Anthropic released Claude Sonnet 5, which the company calls its most agentic Sonnet yet. It can build plans on its own and use tools like browsers and terminals. In benchmarks, Sonnet 5 beats its predecessor, Sonnet 4.6, across the board and closes in on the larger Opus 4.8. On real-world knowledge work tasks, it even edges past Opus 4.8. The model is available now on all Anthropic platforms at an introductory discount, with pricing rising to standard Sonnet rates after August 2026. Anthropic released Claude Sonnet 5. In benchmarks, it closes in on the larger Opus 4.8 and even beats it in some areas. The model is available now at an introductory price. Anthropic calls it the most agentic Sonnet yet: it can build plans, grab tools like browsers and terminals, and work on its own at a level that just months ago only bigger, pricier models could pull off, according to the company. Sonnet 5 is meant to close that gap. Anthropic's published benchmarks show Sonnet 5 beating its predecessor Sonnet 4.6 in every tested category while gaining ground on the pricier Opus 4.8. On agentic coding, Sonnet 5 hits 63.2 percent on SWE-bench Pro, up from 58.1 percent for Sonnet 4.6. Opus 4.8 sits at 69.2 percent. On Terminal-Bench 2.1, Sonnet 5 pulls 80.4 percent versus Sonnet 4.6's 67.0 percent. For multidisciplinary reasoning (Humanity's Last Exam), the model reaches 57.4 percent with tools, nearly matching Opus 4.8 at 57.9 percent. On computer use (OSWorld-Verified), Sonnet 5 posts 81.2 percent compared to 78.5 percent for its predecessor. Ad On the knowledge work benchmark GDPval-AA v2, which tests AI on real-world knowledge tasks , Sonnet 5 actually beats the larger Opus 4.8, scoring 1,618 to Opus's 1,615. Anthropic says feedback from early-access partners told the same story. Sonnet 5 acts far more agentically than previous versions, showing up in things like how it handles search tasks. Ad DEC_D_Incontent-1 Lately, Anthropic has been making news for models it can't ship. The US government is blocking the company's two most capable models, Mythos 5 and Fable 5 , over cybersecurity concerns. That context hangs over the Sonnet 5 launch. Anthropic is clearly eager to get ahead of any similar worries. The model wasn't trained on cybersecurity tasks, the company says, and in tests for risky capabilities like writing software exploits, it scores far below both Opus 4.8 and Mythos 5. Sonnet 5 does score a bit higher than its predecessor on these tasks, though. So Anthropic has switched on cyber safeguards by default. They flag and block risky cyber usage in real time, on par with the protections already in place for Claude Opus 4.7 and 4.8. They're dialed back compared to Fable 5's guardrails, which users complained about almost immediately . Anthropic says it views the overall cybersecurity risk from Sonnet 5 as low. Ad On the safety front, the model does a better job turning down malicious requests and fending off prompt injection attacks than Sonnet 4.6, according to Anthropic. Hallucinations and sycophantic behavior , the tendency to just agree with whatever the user says, are down as well. Anthropic's full safety evaluation is in the Claude Sonnet 5 System Card . Claude Sonnet 5 is live now on all plans. It's the new default for Free and Pro users, and Max, Team, and Enterprise subscribers can access it too. Developers can plug it into Claude Code and the Claude Platform. On the API side, it goes by "claude-sonnet-5" . The training cutoff is January 2026, with a one-million-token context window. Ad DEC_D_Incontent-2 Until August 31, 2026, Anthropic is charging $2 per million input tokens and $10 per million output tokens. After that , prices jump to $3 and $15, which is what previous Sonnet models cost. Ad