Claude Opus 5

Model id claude-opus-5 (released 2026-07-24, Anthropic's fourth Claude 5 model). Positioned as a thoughtful, proactive agent that 'comes close to the frontier intelligence of Claude Fable 5 at half the price' — same $5/$25 per MTok as Opus 4.8 (half of Fable 5's $10/$50); it is the new default model on Claude Max and the strongest model on Claude Pro. familySlug is claude-5, but predecessor is null: it is the first Opus in the Claude 5 family and cross-family predecessors are disallowed (Opus 4 likewise had a null predecessor at the start of Claude 4). Fast mode runs ~2.5x default speed at twice the base price ($10/$50). Per Anthropic's Opus 5 System Card (Table 8.1.A) it is state-of-the-art on FrontierBench v0.1 and GDPval-AA v2, roughly triples the next-best model on ARC-AGI-3, and ~1.5x the next-best on Zapier AutomationBench; it trails GPT-5.6 Sol on DeepSWE and ARC-AGI-2, and Fable 5 on SWE-bench Pro and FrontierCode. HLE headline uses the with-tools configuration (64.7%; no-tools is 56.3%) to stay ordering-consistent with the sibling Sonnet 5 row, whose corpus HLE figure (57.4%) is also with-tools. GDPval-AA v2 is deliberately NOT filed for this model: Opus 5's launch figure (1861) is on the current, re-fitted Artificial Analysis leaderboard (where Fable 5=1747, Opus 4.8=1593), whereas this corpus's gdpval-aa column is an earlier snapshot (Fable 5=1932, Opus 4.8=1890); mixing the two scales would make Opus 5 spuriously lose its GDPval-AA battles to Opus 4.8/Fable 5, contradicting the published SOTA result, so the row is omitted rather than corrupt the Elo. Cyber classifiers engage ~85% less often than Fable 5; flagged cyber/bio requests fall back to Opus 4.8 by default.