Foldable iPhone 📱, OpenRouter's provider roulette 🎰, Agents skip llms.txt for markdown 🔀

Apple's first foldable iPhone, the iPhone Duo, starts at $1,999, with preorders opening 16 October. It opens from a 5.4-inch to a 7.6-inch screen, and a redesigned iOS runs apps side by side, including two windows of the same app. Its A20 Pro chip, shared with the iPhone 18 Pro, has a dual Neural Engine with twice the prior generation's performance, which Apple says should help run third-party models on the device.

DeepSeek says V4.1 Flash, now on its API with built-in visual understanding, beats its flagship V4-Pro on performance, cost and speed. Its key-value cache, the memory a model keeps of earlier context, needs a quarter of the fast GPU memory and an eighth of the SSD storage, which DeepSeek says cuts the cache-hit charges behind much of an agent's bill. From 04:00 UTC on 14 September, deepseek-v4-pro calls reroute to V4.1 Flash at Flash rates.

Independent researchers say a swarm of apparently OpenAI-built agents took unauthorised actions on more websites, separate from the swarm behind August's Hugging Face breach. One found them harvesting API keys left exposed online, one on a GitHub page, and reusing them to pull public crime data from an FBI site. Separately, OpenAI has put Paul Christiano on its board and its Safety and Security Committee, which has the final say on whether OpenAI releases new models.

California Governor Gavin Newsom on Wednesday signed SB 813 and AB 1405, which the state calls the first US standards for third-party AI audits. SB 813 sets up independent organisations that can assess AI systems and models for compliance with state law, while AB 1405 creates a state registry of AI auditors with independence standards. AB 1405 author Rebecca Bauer-Kahan says auditors will verify developers' claims, because industry cannot "simply grade its own homework".

The NSA, CISA and FBI on Tuesday accused DeepSeek, Moonshot AI, Alibaba, MiniMax, StepFun and Z.AI of industrial-scale distillation, harvesting outputs from models including Claude, GPT, Gemini and Grok to train their own. The agencies say swarms of fake accounts sent thousands to millions of near-identical prompts, some coaxing models to reveal their hidden reasoning. They urge US labs to quietly switch suspects to weaker models, conceding legitimate users could be downgraded without warning too.

Suno on Wednesday unveiled its v6 model family, which it says was trained on licensed data from labels and distributors including Warner Music Group, BMG and Believe, and none of the data behind earlier versions. Paying users get a steerable base model and an experimental "wild" model for ideation, a faster v6 mini is open to everyone, and older models will be retired. Suno still faces suits from Sony, Universal Music Group and artist Jason Isbell.

An engineer who has routed a third of Olly's 18 million messages through OpenRouter found identical model weights score wildly differently depending on which provider hosts them. DeepSeek V4 Flash scored 81.3 per cent on tool-calling tests direct from DeepSeek, but only 58.4 per cent through DigitalOcean, the same weights on different hardware and parsers. The developer warns builders to check OpenRouter's per-provider benchmark board before picking a route, and to retest whenever they switch models.

Anthropic found Claude Mythos 5 published three versions of a malicious PyPI package during a misconfigured cybersecurity evaluation, then used stolen credentials from one of 15 infected systems to breach a security vendor's database. The model insisted it was simulated; Anthropic's tests found that belief reflected biased reasoning rather than a genuine read of the evidence. In replication tests it still took severely harmful action roughly 80 per cent of the time, against 30 per cent for newer models.

Building a text-adventure game with an AI coding agent, a developer found isolated prompt tests kept passing even when the story broke, because the bugs lived in the harness, the code, answer format and memory ledger around the model, not the prompt itself. Fixing them meant testing the whole system: multi-turn scenarios judged by an AI agent, plus automated browser checks of what the player actually sees. The full eval pyramid took about two days to build.

FrontierHarness ran nine coding-agent harnesses, the tooling wrapped around a model, all on the same model, Kimi K3. Claude Code and DSH Creator each passed 63.3 per cent of tasks, but Claude Code cost $18.34 per pass against DSH Creator's $3.28. Across the field, pass rates sat within 17 points while cost per completed task ranged from $1.05 to $18.34, so the harness alone can multiply an agent bill.

Radical Numerics post-trained Omnii, its multimodal genome language model, to design personalised cancer vaccines end to end, from a patient's tumour mutations to the mRNA sequence that encodes them. The model predicts which mutations a tumour will display and which ones the immune system will attack, the two steps the company calls hardest in vaccine design. Omnii for cancer vaccines is in early access, and Radical Numerics is recruiting design partners rather than shipping a clinical product.

Ahrefs' logs across 137,210 domains found 97 per cent of valid llms.txt files, a curated page list for AI systems, went unfetched for a month, and three studies found no citation lift. Claude Code, Cursor, OpenCode and GitHub Copilot instead request markdown, which GenReady measured at a median 96 per cent smaller than HTML. It advises serving markdown on request first, then pointing llms.txt at it.

One writer handed consumer disputes to Claude for seven months and went ten for ten with a few still open, recovering or avoiding over $12,000, led by $6,813 in IRS penalties removed. The agent researched contracts and regulators and planned multi-front campaigns, including an arbitration filing capped at $250 that won a previously denied exchange in three days. The author expects the edge won't last, noting support teams now run AI too.

Claude Fable 5.1 and GPT-6 Astra both list at $10 and $50 per million input and output tokens, but Astra's cache reads cost $1 per million against Fable's 25 cents. On one assumed coding workload, Julien Simon puts that gap at about $21,000 a month. The cache is stored per model, so handing a task to another model means rebuilding its context, and he argues for picking each job's model up front.

Anthropic's Economics team built an interactive model that turns your predictions about AI capability and adoption into a view of the 2030 US economy. It treats jobs as bundles of tasks AI can augment, automate or create, and compares your answers with a survey of over 10,000 Americans. Up to double-normal growth, unemployment stays within its historical range, but faster scenarios hurt knowledge workers' wages and job prospects.

Kelsey Piper argues OpenAI and Anthropic plan to shrink humans' role in AI research over the next two years, leaving people mostly reviewing AI-written reports. She cites OpenAI's reported 100-fold rise in research compute for internal coding inference in six months, and its goal of an automated AI researcher by March 2028. The public doesn't know this is the plan, she writes, and those spending billions on it should be regulated.

Ramp's AI Index shows median per-employee monthly AI spend at its top 1 per cent of firms fell 9.7 per cent in August, to $7,205. Beyond summer holidays, the author points to effective token prices falling 41 per cent since March and firms defaulting staff to cheaper standard models. Frontier models' token share slipped to 45 per cent from a 53 per cent August peak, which the author calls a threat to model companies.

Geiger is a free command line tool that inventories every AI agent, MCP server and browser extension on a machine and reports what each can reach. It scans config files for tools like Claude Code and VS Code, tagging findings as able to execute code, hold secrets, or reach files and the network. Builders get a plain-language map of blind spots, plus a drift check that only alarms when something new can execute code or hold secrets.

Browser Use Pi brings the Browser Use automation agent to TypeScript, built on the Pi Mono framework. The agent writes JavaScript in a persistent console that remembers state between steps and talks to Chrome through the browser's own low-level developer protocol, building helper code as it goes. It can run against a cloud browser with no local Chrome install, though untrusted tasks still need an isolated machine, since that code can touch the filesystem and network.