Skip to content- Watching for: Independent benchmark reproduction of the coding/agentic and ARC-AGI 3 claims; whether Claude Code moves its default model to Opus 5; head-to-head against Fable 5, GPT-5.6 Sol, and Kimi K3 on real coding tasks; any change to the weekly-cap/metering terms; deprecation timeline for Opus 4.8.
Notes
- 2026-07-24: Anthropic released Claude Opus 5 on 2026-07-24, new flagship, API id
claude-opus-5, priced $5/M input and $25/M output (same as Opus 4.8), available in the Claude API, on Claude.ai, and in Claude Code, with a fast mode at 2.5x speed for 2x base price. Vendor benchmarks (unreproduced): SOTA on Frontier-Bench and GDPval-AA coding, more than 2x Opus 4.8 on Frontier-Bench v0.1, ~3x next-best on ARC-AGI 3, OSWorld 2.0 above Fable 5 at lower cost. System card: most aligned Anthropic model to date by automated behavioral audit, still behind Mythos 5 on cyber exploitation and biology, safeguards similar to Opus 4.8. Covered 2026-07-24 Top stories (lead, confirmed). 2026-07-24/25 first third-party measurement: Artificial Analysis Intelligence Index v4.1 (nine evals incl. GDPval-AA v2, Terminal-Bench v2.1, SciCode, HLE, GPQA Diamond, AA-Omniscience) ranks Opus 5 at adaptive reasoning/max effort first at 61, ahead of Fable 5 (60) and GPT-5.6 Sol max (59); Opus 5 also holds second (xhigh, 60) and fifth (high, 59), so the ranking depends on effort configuration (HN 49040741). Anthropic's Boris Cherny states Opus 5 is the least prompt-injectable Anthropic model so far, referencing prompt-injection evals and red-teaming on system card p.73, with no numbers (relayed by Simon Willison 2026-07-25). Covered 2026-07-25 Watchlist follow-ups (developing). See [[entities]] Anthropic.
Stories