Fable 5 returns worldwide 🤖, Claude Code fingerprints you 🔑, a 26B model on a bare CPU 🧑‍💻

Anthropic will restore global access to Claude Fable 5 on Wednesday, after the Commerce Department lifted the export controls it imposed on 12 June over the model's cyber capabilities. A licence is no longer required to export Fable or Mythos 5. Anthropic's own testing confirmed Opus 4.8, GPT-5.5 and Kimi K2.7 could find the same vulnerabilities, and it has agreed a new jailbreak-severity framework with Amazon, Microsoft and Google.

Anthropic's new midsize model lists cheaper than Opus 4.8, but Artificial Analysis found it costs more to run: Sonnet 5 generates nearly double the output tokens per task, landing at $2.29 against Opus's $1.80, above even Fable 5 across the full benchmark. The same day, Anthropic also shipped Claude Science, a workbench that folds a researcher's whole toolchain into one auditable app.

The Nvidia challenger has taped out its transformer-specialised inference chip at TSMC and booked a billion dollars in orders for full 'frontier inference clusters' now in customer testing. Founded in 2022 by two Harvard dropouts who nearly ran out of cash, Etched has raised $800 million at a $5 billion valuation. Its angel list reads like an AI hall of fame: Karpathy, Hinton, Fei-Fei Li, Mensch and Scott Wu.

The food-delivery giant's LongCat-2.0 is a 1.6-trillion-parameter mixture-of-experts model with around 48 billion active per token, available openly with a one-million-token context. What makes it notable is the hardware: the full training run and serving happen on domestic AI ASIC superpods, across 35 trillion tokens with no loss-spike rollbacks. It plugs straight into Claude Code, OpenClaw and Hermes for repository-level edits and agentic work.

A trafficked worker in a Myanmar scam compound impersonated a Singaporean woman and targeted 50,000 victims across 17 countries in a single month, supervisors patrolling the desks with electric batons. An AP and Frontline investigation places American AI models, Starlink dishes and ISPs all along the supply chain that industrialised this fraud. US watchdogs say the tech firms could do more but lack the incentives to act.

The open-source game engine behind Slay the Spire 2 will amend its contributor guidelines to forbid AI-authored code, pull requests opened by agents, and AI-generated text in human discussion. Maintainers called the flood of AI slop pull requests draining and demoralising, and hard to justify reviewing when the feedback only trains a machine. Their line: heavy AI users often cannot understand their own code well enough to fix it.

With DDR5 prices at record highs, Meta is pulling DDR4 modules out of decommissioned servers and reattaching them to new AMD Turin machines through its own Vistara ASIC, a CXL 2.0 memory expander over PCIe 5.0. Each server pairs 768GB of fast local DDR5 with 256GB of recovered DDR4, and Linux quietly migrates cold pages to the slower tier. It is a hardware admission that the memory shortage is biting.

Amazon is the first hyperscaler to formalise a Forward Deployed Engineering unit, seeding it with thousands of engineers who embed in pods of five or six inside a customer to accelerate AI rollouts. The Palantir-coined model has spread fast: OpenAI and Anthropic stood up their own FDE arms earlier this year with banks and consultancies. AWS frames it as pulling scattered capability into one business with a shared rubric.

A developer decompiling Claude Code 2.1.196 found a function that rewrites the date string in the system prompt, silently switching the apostrophe in Today's and the date separator from dashes to slashes. The variants act as a steganographic marker on outgoing requests, with logic that also checks for Chinese timezones. It is a sharp reminder to audit what your coding harness injects before you hand it your filesystem and shell.

Chasing how fast a model can think on hardware you own, a builder ran Gemma-4-26B, a mixture of experts firing 3.8 billion parameters per token, on an i9 with plain DDR5, hitting 40 tokens a second single-stream and 124 batched, losslessly. The counterintuitive lesson: bytes read per token, not parameter count, set the speed. The experts are only 16% of reads, so you compress the 262K-token vocabulary head instead.

Fergus Finn takes the simplest CUDA program, a vector addition, and follows it all the way down, from the source you write to what the GPU actually does to run it. It is the kind of first-principles walkthrough that turns 'I call a kernel and it goes fast' into a real mental model of blocks, threads and memory. Worth it if you write GPU code and want the layer beneath the abstraction.

Rather than guessing from ChatGPT's answers, Suganthan Mohanadasan inspected the network traffic behind them to work out which sources the model actually fetches, cites or quietly ignores. The write-up surfaces the hidden request fields that shape that selection, a rare look under the hood for anyone trying to get their content cited by an assistant. A practical read for builders thinking about how retrieval and grounding choose winners.

Meta's second-generation Brain2Qwerty reads magnetoencephalography recordings and decodes full sentences in real time from non-invasive brain signals, reaching 61% word accuracy overall and 78% for its best participant, up from 8% for prior non-invasive methods. The team trained it end-to-end on 22,000 sentences and has published the full training code. Accuracy scales log-linearly with data, hinting the gap to surgical implants could keep closing.

Ethan Mollick argues the useful frame is no longer the chat box but the autonomous agent, and the capability curve is bending better than exponential. He cites Epoch's finding that Opus 4.7 worked alone for 14 hours to build software worth two to seventeen weeks of human engineering, for $251 in tokens. His own tests had Fable running nine hours on projects a team would need over a week to finish.

Lenny Pruss makes the case that AI inverts decades of software logic: value came from owning workflows and features, but agents do not navigate a GUI, they read your API spec and compose capabilities. The reliable product for an agent is simply the best product for a developer, so every software company now has to think like a dev-tools company. He reaches for HashiCorp's Tao: collapse complexity into the smallest powerful abstraction.

Ben Bajarin's constraint map treats announced gigawatts and GPU orders as intent, not billable compute, because a purchase order becomes revenue only when every layer lands at the same site in the same window. Chips set the pace early, but the binding constraint has moved deeper: cooling loops, power equipment, permitting, networking and the crews that certify a rack. A campus can hold its GPUs and still earn nothing.

Rich Mironov warns founders against confusing code acceleration with revenue acceleration. If most products fail and you now ship 100 where you shipped 10, budgets and markets have not grown to match, so the same four winners emerge and the success rate collapses from four-in-ten to four-in-a-hundred. Cheaper production just moves the bottleneck back to the hard part: finding real buyers with real pain and getting to market.

This essay traces how 1990s computing forced you to learn the machine, editing autoexec.bat and setting drive jumpers, because it would not run until you did. The resistance was the medium: you only truly know the things you can lose to. AI is the most accommodating tool ever built, rearranging itself around your sentence and never setting terms, and a machine that cannot push back is one you can only use.

Cognition's new multi-model harness runs a frontier model and a cheaper 'sidekick' agent in parallel, with the main agent delegating routine work and keeping the plan, ambiguity calls and final review to itself. Cognition says the setup holds Fable-5-level quality at 35% lower cost on its FrontierCode benchmark, which grades both correctness and mergeability. Dynamic mid-session routing decides task by task who does what. It is in preview now.

Assembled has open-sourced 143.dev, the internal system it built to run coding agents as shared team infrastructure rather than a tangle of personal setups. It runs Codex, Claude Code and OpenCode in gVisor sandboxes so you can mix harnesses to trade cost against intelligence, and wires in GitHub, Linear, Sentry, Slack and PagerDuty with easy previews. Inspired by Stripe's Minions and Ramp's Inspect, it is MIT-licensed and self-hostable.

Google's Agent Development Kit for Go reached 2.0, aimed at building reliable multi-agent applications with more structure than a bare prompt loop. The release adds a graph-based workflow engine, built-in human-in-the-loop approval gates, and dynamic orchestration that routes work at runtime. For Go teams standing up production agents, it is a first-party framework with the control points that ad hoc harnesses usually leave you to build yourself.

Meta has open-sourced Astryx, a fully customizable design system written in TypeScript and built to be agent-ready, so coding agents can read and compose its components rather than only humans clicking through them. It is trending near the top of GitHub, adding hundreds of stars a day. As agents start consuming interfaces directly, a component library they can parse becomes a real foundation rather than a nicety.