According to Technews, Anthropic released Claude Opus 5 last Friday, marketed as delivering "接近頂級模型 Claude Fable 5 的前沿智慧" (near-frontier intelligence close to Fable 5) at half the price (E1). Inside.com.tw reports that Anthropic's own announcement is more cautious: the company states Opus 5 is "接近" (close to) Fable 5's capability, but still recommends Fable 5 for the most complex projects (E11). This positions Opus 5 as a mid-tier model rather than a direct Fable 5 replacement.
Per-million-token pricing, as reported by Technews, shows a clear tier structure:
| Model | Input ($/M tokens) | Output ($/M tokens) |
|---|---|---|
| Claude Fable 5 | 10 | 50 |
| Claude Opus 5 | 5 | 25 |
| GPT-5.6 Sol | 5 | 30 |
(Source: Technews, E2)
Technews also cites a weighted-average cost-per-task figure from an intelligence-index benchmark, which ranks actual task cost differently than list price:
| Model | Cost per task ($) |
|---|---|
| Claude Fable 5 | 2.75 |
| Claude Opus 5 | 2.03 |
| GPT-5.6 Sol | 1.04 |
| Kimi K3 | 0.95 |
(Source: Technews, E3)
Inside.com.tw adds that on the hardest benchmark tier, CursorBench 3.2, Opus 5 trails Fable 5 by only 0.5% while costing half as much (E9). Anthropic also offers a "fast mode" that costs twice the standard price in exchange for 2.5x processing speed (E12), giving users a second cost/speed lever beyond the base pricing.
Technews reports four specific benchmark results for Opus 5: on Frontier-Bench v0.1 (software engineering), Opus 5 outperforms all rivals with more than double the performance of predecessor Opus 4.8, at lower per-task cost; on ARC-AGI 3 (novel problem-solving), Opus 5's score is three times that of the next-best model; on Zapier AutomationBench (business automation), its end-to-end completion rate is 1.5 times the next-best model at equivalent cost; and on OSWorld 2.0 (computer-use tasks), it surpasses all models at equal cost, exceeding even Fable 5's best score (E4).
Inside.com.tw's benchmark citations overlap on some points but add detail: on OSWorld 2.0, Opus 5 beats Fable 5's performance at one-third of Fable 5's cost (E9). On safety-flavored evaluation, Anthropic calls Opus 5 the "most consistent Opus model" and the hardest version to induce into misuse, but the company acknowledges Opus 5 still lags Mythos 5 on cybersecurity tasks — described as "close to Mythos 5 on vulnerability discovery, but far behind on exploitation" (E14).
According to Inside.com.tw, Anthropic added an "effort" control that lets users choose among four reasoning-intensity levels — low, medium, high, and xhigh — directly determining how much inference cost the model is willing to spend on a given question (E10). Technews reports a separate API-level feature called "Auto-fallback": when a user's request is judged too sensitive for Opus 5 or Fable 5 and triggers a safety classifier, the system no longer blocks the request outright but automatically downgrades it, so the workflow is not interrupted (E7).
Technews reports that Anthropic engineer Thariq Shihipar noted the company removed as much as 80% of Claude Code's system prompts in this release, and recommends developers use the "claude doctor" command to auto-optimize their existing prompts (E5). On adversarial testing, Technews cites OSS-Fuzz results showing Opus 5's open-source vulnerability discovery rate at 80%, roughly on par with security-focused Mythos 5's 79.4%. However, Opus 5's ability to "weaponize" a found vulnerability is far more restricted — a 4-out-of-14 success rate compared with 13-out-of-14 for Mythos 5 (E6).
Inside.com.tw reports that Anthropic released Opus 5 on July 24, 2026, across Claude.ai, Claude Code, Claude Cowork, and the API simultaneously, making it the new default model for the Claude Max subscription plan, with pricing kept identical to predecessor Opus 4.8 at $5/$25 per million tokens (E8). The same report notes Opus 5 is the fourth model Anthropic has released within two months — following Mythos 5, Fable 5, and Sonnet 5 in June — and that this release cadence coincides with Anthropic's preparations for an initial public offering (IPO) this year (E13).
The pricing and benchmark data point to a consistent story across both sources: Opus 5 is built to undercut Fable 5's list price by half (E1, E2) while closing most of the performance gap on specific tests — within 0.5% on CursorBench 3.2 (E9) and ahead of Fable 5 on OSWorld 2.0 at a fraction of the cost (E4, E9). Yet Anthropic's own framing keeps Opus 5 in a secondary tier, still recommending Fable 5 for the most complex work (E11) and acknowledging it trails Mythos 5 specifically on cybersecurity exploitation, 4/14 versus 13/14 (E6, E14). That combination — a cheaper, benchmark-competitive model launched as the fourth release in two months (E13) on the same day pricing was held flat versus Opus 4.8 (E8) — lands directly in the window Anthropic has designated for IPO preparation this year (E13).
According to NVIDIA's newsroom, NVIDIA has expanded its NVIDIA Agent Toolkit for engineering by adding NVIDIA PhysicsNeMo and three new CUDA-X libraries — cuISS, cuDSS and cuEST — as agent-ready tools. NVIDIA states partners including Cadence, Siemens, Samsung, Silvaco, Keysight, Synopsys and TSMC have reported speedups ranging from 10x to 50x using the new tools.
According to NVIDIA's official blog, the company is deploying its 88-core Vera CPU across the EDA workflows used to design its next-generation chips, and is working with Cadence and Synopsys on tools that showed up to 1.5x higher performance on selected workloads.
According to NVIDIA, SK Group and NVIDIA unveiled a $500-billion-plus partnership, launching a 2-gigawatt AI factory using SK hynix HBM4 in 2027.