Anthropic Releases Claude Opus 5 — Near-Fable 5 Performance at Half the Price
On July 24, Anthropic released its latest LLM, **Claude Opus 5**. The model delivers intelligence approaching the higher-tier **Claude Fable 5** at half the price. It has earned strong reviews in programming and knowledge work, and reportedly ships with a parameter for tuning reasoning depth plus an automatic fallback feature tied to its safety classifiers. Its cybersecurity safeguards also appear to have been adjusted in a more permissive direction, presumably due to operational concerns — a clear eye toward operational balance.
Shortly after, **Claude Code v2.1.219** was released, naming Opus 5 as the new default Opus model. With support for the 1M context and fast mode, it marks a real transition point for users moving off the Opus 4 line.
Fable 5 "Creditization" Lands — Permanently Free up to 50% for Max Tiers, $100 Credit for Pro Tiers
Also in July, Fable 5's usage terms reached their final form. Seats on Max / Team Premium / Enterprise Premium get **permanent free usage up to 50% of the weekly cap**. Pro / Team Standard, meanwhile, receive a **$100 credit** on top of their in-plan allowance, after which usage shifts to pay-as-you-go at $10 / $50. The long-debated two-tier "subscription + metered" structure is now official, and it could well shape how competitors design their pricing.
Reid Hoffman and Marc Pincus Found New AI Lab "Prentis" — $100M Raise Under Discussion
LinkedIn co-founder **Reid Hoffman** and Zynga co-founder **Marc Pincus** have co-founded a new AI lab, **Prentis**. According to TechCrunch, it is in talks to raise $100M. The lab's thesis is that "automating everyday computer work will become AI's biggest use case, edging out coding," and it is aiming squarely at the next wave of the agent space. Hoffman is also known for his involvement with Inflection AI, and this restart by someone with deep big-platform connections is drawing plenty of attention.
Hermes AI Agent Abused in Attack on Thailand's Finance Ministry — A New Threat for the Agent Era
According to BleepingComputer, the **Hermes AI agent** was used to automate an attack on Thailand's Ministry of Finance. Agentic AI boosts productivity, but it also carries the risk of being abused by attackers as "a tool that automates procedures." Defenders will need to wield agents just as actively, and how to close that asymmetry looks set to become a key challenge.
Debian Debates a Ban on LLM Contributions — Redrawing the Rules of OSS Communities
A **General Resolution (GR)** has been put forward in the Debian Project to ban contributions of LLM-generated material to Debian. The main drivers are unclear copyright provenance and licensing concerns, and the outcome could ripple out to other major OSS communities. Debian is known for its strict Social Contract, and this decision could become a symbolic moment for where the free software world stands.
Gemma 4 26B A4B Runs on an iPhone 17 Pro — Via "Model Paging"
A report that **Gemma 4 26B A4B** runs on an iPhone 17 Pro has been making waves in the LocalLLaMA community. It works through Noema's "Overfit" system, which uses a **model paging** approach: non-expert weights are kept in RAM while expert weights are read from the SSD. At 34.4 tok/s prefill and 3.5 tok/s decode, speed involves some compromise, but the verdict is that it's practical for cases where answer accuracy matters. The technique could also be applied to low-end MacBooks and the like — a step that extends the reach of local LLMs.
Other Moves
A Canadian member of Parliament read what looked like an LLM response aloud in a floor speech** — telltale signs were spotted and it went viral on social media (Ars Technica).
Overture Maps releases a Cross-Theme Knowledge Graph prototype** — a map knowledge base for grounding LLMs (Overture Maps).
Networking expert Glenn Fiedler warns game developers** — we're entering an era where AI can find vulnerabilities in game code (GamesBeat).
Validating demand for a "product that doesn't exist yet" with video-generation AI** — a hands-on article on how to split work across Runway / Kling AI / Sora (Qiita).
---