Cheaper Models, Whole-Body Robots, and Amodei Disagrees on Open Weights
2026-07-30
Anthropic released Claude Opus 5, matching Fable 5's coding and knowledge-work performance at half the price, while its CEO published a position on open-weights models that split the industry. Black Forest Labs launched FLUX 3, a multimodal model spanning video, audio, images, and robotic action from a single architecture. Google's Gemini Robotics 2 brought whole-body control to humanoid robots, and OpenAI cut prices across the GPT-5.6 line while adding a Fast mode to GPT-5.6 Sol.
News
Anthropic released Claude Opus 5, a model that matches Claude Fable 5 on coding and knowledge work benchmarks at Opus-tier pricing ($5/$25 per million tokens). It scored within 0.5% of Fable 5 on CursorBench 3.2 at max effort and set new state-of-the-art results on FrontierBench, ARC-AGI 3, and OSWorld 2.0. On Frontier-Bench v0.1 it more than doubled Opus 4.8's score. Its cyber capabilities improved substantially over Opus 4.8 — nearly matching Mythos 5 at vulnerability discovery — but it was deliberately not trained on cyber tasks and remains behind Mythos 5 on exploit development. It is Anthropic's most aligned model yet on automated behavioral audits, with the lowest cooperation-with-misuse score of any model tested.
Google DeepMind introduced Gemini Robotics 2, a family of three models for robot control. Gemini Robotics 2 is a vision-language-action model that drives full humanoids from feet to fingertips and handles dexterous manipulation with both hands and grippers; Gemini Robotics ER 2 is an embodied-reasoning model for multi-step planning and multi-robot teamwork; and Gemini Robotics On-Device 2 runs locally and adapts to a new robot body in a few hours. The ER 2 model is live on Google AI Studio, with the other two available to early-access partners.
OpenAI cut prices across the GPT-5.6 line, making Luna 80% cheaper and Terra 20% cheaper, and adding a Fast mode to GPT-5.6 Sol that runs up to 2.5x faster at twice the price. OpenAI positions Luna as delivering performance comparable to models that were frontier-class a year ago at roughly 6 cents on the dollar per task, at nearly nine times the speed.
Anthropic CEO Dario Amodei published his position on open-weights models, stating the company has never advocated for a ban and calling open-weight models without dangerous capabilities "a public good." His three preferred interventions are chip export controls, crackdowns on industrial-scale distillation, and mandatory pre-release safety testing for all sufficiently capable models. He explicitly disagreed with an open letter signed by Nvidia, Microsoft, Meta, and 77 others urging support for open-weight models, arguing broad access may help attackers more than defenders, especially in biology. The letter itself gained 624 points on HN.
Black Forest Labs launched FLUX 3, a multimodal foundation model trained jointly on images, video, and audio within a single architecture — then extended to robotic action prediction. It generates video up to 20 seconds with native synchronized audio, supports text/image/video-to-video and keyframe control, and already runs on Audi production lines through FLUX-mimic, handling soft-body manipulation tasks conventional robotics cannot. In preliminary evaluations, FLUX 3 was preferred over Runway Gen-4.5 (77%), Kling v3 Pro (60%), and Luma Ray 3.2 (93%). An open-weight FLUX 3 Dev backbone is planned for later this year. Video is in gated early access; image generation and pricing are not yet available.
Articles
Anthropic removed over 80% of Claude Code's system prompt for Claude 5-generation models with no measurable loss on coding evaluations. The engineering team describes a shift from giving Claude rules and examples to designing interfaces and using progressive disclosure — loading the right context at the right time instead of front-loading everything. A new claude doctor command automatically rightsizes skills and CLAUDE.md files.
Six weeks after the Bun Rust rewrite announcement, the project has no release tag and 2,475 open robobun PRs remain unmerged. The original $165K API cost omitted CI/CD expenses and ongoing Claude usage, pushing the estimated total toward $800K. The Rust-powered Bun has been shipping in Claude Code since June 17, but the open-source release is delayed waiting on a Node.js compatibility milestone. Anthropic employees are directly contributing Rust commits. Zig creator Andrew Kelley responded that the rewrite "had nothing to do with language features" and everything to do with diverging engineering values.
A US citizen faces federal felony charges for giving border agents a GrapheneOS duress code that wiped his phone during a customs search at Atlanta's airport — the first known prosecution of its kind. The government is using a rarely-invoked statute criminalizing destruction of property to prevent seizure. The defendant's lawyers argue he was targeted over activism against the Cop City training facility. The EFF notes they have never seen the law applied this way.
AI companies are bulk-buying physical books — including rare and out-of-print titles — and destroying them through destructive scanning to obtain pre-2022 training data free of AI-generated text contamination. ISBNdb now markets orders of up to one million books to AI developers with NDAs as a feature. Booksellers report a surge in bulk purchases since April, often targeting foreign-language and uncommon titles whose scarcity is not tracked at the copy level. A 2025 federal ruling treating Anthropic's destructive scanning as fair use provides the legal scaffolding.
Tools
GitHub launched stacked pull requests in public preview, breaking large changes into a chain of small, dependent PRs reviewable independently and mergeable in one click. Includes a gh stack CLI extension with full branch management, cascading rebases, and a terminal UI for restructuring stacks. Battle-tested by Next.js, TED, and jQuery. (GitHub CLI / Go)
PGSimCity is an explorable 3D city visualization of PostgreSQL internals, built with three.js and TypeScript. Every building maps to a real mechanism — shared_buffers plaza, WAL district, autovacuum trucks, a query lab — with semantic color coding. Built with 3.86 billion tokens of AI-assisted coding. Press G to walk the city at eye level. (Browser / TypeScript)
BannerBye is a browser extension that kills cookie banners before they load, using GPC (Global Privacy Control), IAB TCF v2.2, and CMP-specific API signals to tell sites consent was already refused. Falls back to auto-clicking "Reject all" buttons with a remotely-updated ruleset. Open source (MIT), available on Chrome, Firefox, and Safari. (Browser extension / TypeScript)
A security camera shipped a live GitHub admin token in its login page, embedded in the HTML of the device's web interface. The token granted full repository access. The manufacturer had apparently committed credentials to source control and deployed them to production firmware. (Hardware security)
Worth a Look
HTMX 4.0 released exclusively as a physical Game Boy cartridge — a real Mario-inspired platformer with four levels, three biomes, and a final boss battle against "Warren Buffering" that unlocks the 4.0 source code. 1,000 cartridges manufactured, 150 given away at Big Sky Dev Con. Built with GBStudio and hand-crafted pixel art.
Kill the Cookie Banner, a campaign by European civil society organizations, is pushing back against Google and tracking industry lobbying that convinced EU member states including Germany and France to strip a proposal from the Digital Omnibus that would have replaced cookie banners with browser-level privacy signals. The solution already exists in California law.
Google says Android may soon restrict on-device ADB connections to only the wlan0 interface, breaking loopback ADB and tools like Shizuku, Termux, and aShell. A detailed technical rebuttal argues a malicious app cannot establish an ADB connection without explicit human action at multiple steps — the real risk is minimal while the ecosystem of tools affected is substantial.
Chrome registers a global keyboard shortcut for a Gemini popup window, overriding existing shortcuts without asking. Users found Gemini appearing on Ctrl+Space or similar chords, with the shortcut buried in chrome://settings rather than chrome://extensions/shortcuts. The change shipped server-side and affected users mid-session.