AI News — July 25, 2026: Opus 5 Tops Leaderboard at Half Fable's Cost, 25 Firms Fight Open-Weight Curbs

Anthropic released Claude Opus 5, which topped the Artificial Analysis Intelligence Index at 61, with Fable 5 and GPT-5.6 Sol within a point or two, at roughly half Fable's cost and without Fable's 30-day data retention requirement. A coalition of 25 companies including Nvidia, Microsoft, Meta, Hugging Face, and Mistral signed an open letter warning against premature restrictions on open-weight models as the Trump administration weighs sanctions against Chinese AI companies. The Guardian published an op-ed arguing OpenAI's 'rogue agent' narrative fits a pattern of using dangerous-AI framing to signal power to investors and regulators.

Good morning. Anthropic dropped Opus 5 on a Friday with none of the doom-laden marketing that accompanied Fable 5, and the model promptly took the top spot on the Artificial Analysis leaderboard. Meanwhile, the open-weights fight escalated: 25 tech giants signed a joint letter against Chinese model restrictions, and skepticism is finally catching up to OpenAI’s Hugging Face incident narrative. Opus 5 arrives, quietly beats everything. Anthropic released Claude Opus 5 https://www.anthropic.com/news/claude-opus-5 , positioning it between Claude Pro and Fable 5 at roughly half Fable’s cost, with adjustable “effort” settings and state-of-the-art scores on Frontier-Bench, ARC-AGI 3, and OSWorld 2.0. The most consequential detail for enterprise buyers, flagged repeatedly in the HN discussion https://news.ycombinator.com/item?id=49038433 : unlike Fable 5, Opus 5 carries no 30-day data retention requirement. The Verge notes https://www.theverge.com/ai-artificial-intelligence/970105/claude-opus-5-announced-anthropic-ai-model-release Anthropic shared the model with government partners before release, and TechCrunch reports https://techcrunch.com/2026/07/24/anthropic-launches-opus-5/ safety classifiers now trigger 85% less often than on Fable 5 — plus a new “Automatic Fallbacks” beta that routes flagged requests to a weaker model instead of erroring out. The leaderboard win comes with an asterisk. Opus 5 tops the Artificial Analysis Intelligence Index https://artificialanalysis.ai/models at 61, with Fable 5 and GPT-5.6 Sol clustered within a point or two. The HN thread https://news.ycombinator.com/item?id=49040741 is unimpressed: GPT-5.6 Sol matches Opus 5’s score at roughly half the price, and Kimi K3 does similarly. One commenter joked that Anthropic’s own contradictory messaging — claiming Opus 5 isn’t “more capable overall” than Fable 5 while listing benchmarks where it wins — is itself becoming a running theme. Developers testing image-to-HTML conversion reported Opus 5 actually beating Fable 5 on visual tasks, which raises the obvious question of what Fable 5 is still for. Big tech lines up against open-weight restrictions. A coalition of 25 companies including Nvidia, Microsoft, Meta, Hugging Face, and Mistral signed an open letter https://www.cnbc.com/2026/07/24/nvidia-microsoft-meta-open-weight-ai-models.html warning against “premature restrictions” on open-weight models, as the Trump administration weighs sanctions against Chinese AI companies over the Fable distillation claims we covered earlier this week. OpenAI initially wasn’t on the list but ultimately signed; Google and Amazon didn’t. TechCrunch’s coverage https://techcrunch.com/2026/07/24/as-us-weighs-response-to-chinese-ai-industry-urges-against-broad-open-weight-restrictions/ quotes Replit’s CEO: “Banning Chinese open models is as good as banning open models in general.” HN commenters read the signatory list cynically — Nvidia wants GPU sales, Meta has idle compute, Microsoft has no strong in-house model — but the underlying point about enforceability is real. Separately, startup founders made a parallel case to Politico https://www.politico.com/news/2026/07/22/startup-founders-urge-trump-not-to-shut-off-chinese-open-weight-ai-01008992 , framing restrictions as regulatory capture. The Hugging Face story gets a skeptical rereading. A Guardian op-ed https://www.theguardian.com/technology/2026/jul/24/openai-rogue-hacker argues OpenAI’s “rogue agent” narrative fits a pattern going back to GPT-2 of using dangerous-AI framing to signal power to investors and regulators. The HN thread https://news.ycombinator.com/item?id=49038060 is split three ways: some agree the marketing incentives are obvious, some point out the piece offers skepticism without evidence, and others argue that reflexively dismissing capability claims as PR is its own form of denial. Between this and the “OpenAI misconfigured a sandbox” reframe from earlier in the week, the incident’s original “AI escaped containment” framing has taken a beating. Prentis raises at a unicorn valuation. Reid Hoffman and Mark Pincus’s new AI lab Prentis is in talks to raise $100M at a $1B valuation https://techcrunch.com/2026/07/24/prentis-new-ai-lab-co-founded-by-reid-hoffman-mark-pincus-in-talks-to-raise-100m/ , just three months after launching. The focus is computer-use models for office workflows, and the company claims its Hive-32B beats GPT-5.4 and Claude Opus 4.6 on key benchmarks at a tenth the cost. Prentis says it has $50M in signed contracts and expects a $75M run rate by Q3 — real revenue in a segment now packed with Anthropic, OpenAI, and Thinking Machines. That’s the week. Weekend homework: figure out whether you actually need Opus 5 or whether GPT-5.6 Sol at half the price is fine.