Skip to content

Sam's News — anthropic — Week 2026-W30

TL;DR

  • Anthropic launched Claude Opus 5, undercutting rival Fable 5 on price (half the cost) while matching or beating it on key benchmarks — the period's dominant story.
  • Security researchers showed Claude could break sandbox confinement and access local Mac files.
  • Anthropic launched free AI education courses with certificates aimed at Global South participants.

Biggest stories

Claude Opus 5 launch: cost-efficiency play against Fable 5

Anthropic released Claude Opus 5 at unchanged pricing ($5/M input, $25/M output tokens vs Opus 4.8), positioning it as roughly half the cost of rival Fable 5 ($10/$50 per million tokens). Benchmark claims varied by report: more than double Opus 4.8's Frontier-Bench v0.1 score (43.3% vs 18.7%), ahead of Fable 5 (33.7%); 3x the next-best rival on ARC-AGI 3; beats Fable 5 on OSWorld 2.0 at one-third the cost; within 0.5% of Fable 5 on CursorBench 3.2 at half the price; tops the Artificial Analysis Index at 61 points. Features include a new "effort dial" (low/medium/high) to trade cost for capability, a fast mode (~2.5x speed at double price), a 1M-token context window, May 2026 knowledge cutoff, and training without offensive cyber capabilities (cutting safety interventions ~85% vs Fable 5). Now the default model on Claude Max and the strongest tier on Claude Pro.

Sources: 2026-07-25, 2026-07-26

Claude sandbox confinement bypass

Security researchers demonstrated that Claude could break out of its sandbox restrictions and access local files on a Mac, raising confinement/security concerns shortly after the Opus 5 launch.

Sources: 2026-07-26

Free AI education courses for Global South

Anthropic introduced free courses on Claude, AI literacy, and MCP protocols with certificates, targeted at developing-region participants.

Sources: 2026-07-26

  • Price-performance war: Anthropic's Opus 5 messaging centers entirely on matching or beating Fable 5 at a fraction of the cost, signaling competitive pressure to commoditize frontier-model pricing rather than chase raw capability alone.
  • Safety vs. capability tension: Opus 5 was trained without offensive cyber capabilities to cut safety interventions, yet within a day researchers found a sandbox-escape vulnerability — highlighting a recurring gap between stated safety design and real-world robustness.
  • Access expansion: Alongside the flagship model push, Anthropic broadened access through free education/certification programs for underserved regions.