Anthropic opens Claude Code to custom mods written in TypeScript
Claude Code just got a mod system. You can now write small TypeScript plugins that hook into the agent loop, change how it behaves, and draw custom UI. Think of it like browser extensions, but for your AI coding tool.
Three built-in mods show what’s possible:
- Blast Radius intercepts risky commands like rm -rf or git reset –hard, shows you what would be affected, and asks Proceed or Cancel before anything runs.
- Token Weather shows a live bar above your prompt with how full your context window is, plus a sparkline of your last 12 turns so you know when to start a fresh session.
- Replay Theater records every file edit Claude makes, then lets you step through each diff one at a time with /replay.
A mod is just a TypeScript file with hooks. You can write one yourself, or ask Claude to build it for you. Install with /plugin in the CLI or desktop app.
One heads-up: mods run with full access to your machine, same as Claude Code itself. Only install from sources you trust.
New open-source prompt skill rewrites stiff AI-generated Japanese into natural text
You know that stiff, robotic feel AI-generated Japanese text has? Someone built a fix for it.
yomiyasu is an agent skill that rewrites AI-generated Japanese into natural, human-readable text. The old approach was just banning certain words, which only swapped one awkward word for another. yomiyasu goes deeper.
It targets two root causes:
- Broken sentence structure (who did what to whom gets lost in AI output) — yomiyasu restores the subject, verb, and object so each sentence stands on its own
- Metaphor verbs like “silently breaks” or “dissolves time” get replaced with precise, literal descriptions of what actually happens
It also ships with a Python linter that scores your text for AI-ness, flagging overuse of bold, bullet points, and vague phrasing.
You can drop it into Claude Code, Codex, or Cursor right now. Install with npx skills add nanaism/yomiyasu, then just paste your draft and say “make this readable.”
OpenAI upgrades Codex Security Cloud to scan entire GitHub repos automatically
Codex Security Cloud just got a big upgrade. Think of it as a security guard that never sleeps, watching your GitHub repo around the clock, even when your laptop is shut.
Here is what it actually does for you:
- Scans your entire GitHub repo and flags vulnerabilities
- Watches every new commit and reviews it automatically
- Removes duplicate alerts so you are not drowning in noise
- Prepares a ready-to-review fix, not just a warning
The big deal here is Daybreak Blue, OpenAI’s defensive security AI tier, which is now bundled in by default. No separate application needed. It runs on a model with a 1,050,000-token context window, meaning it can read your entire codebase at once, not just snippets.
Previously, accessing these stronger cyber models required a separate approval process. Now you just get it. Install the plugin in Codex desktop or web, connect your GitHub repo, and let it run.
Amazon and OpenAI build their own decision models, a category TypeSafe AI launched first
Amazon released Strands Decider 2B, an open-source model that picks from a fixed set of options instead of generating text, built on a small Qwen3.5 base. It started as an internal side project by AWS engineer Marc Brooker, who built it after seeing Jev, a similar model from startup TypeSafe AI. It runs locally, returns answers in under 100 milliseconds, and is fully open source with all training data and scripts included.
Marc’s version briefly topped the JevBench leaderboard before Amazon turned it into an official product. OpenAI shipped a competing Decisions API at DevDay the day before, and Cloudflare followed with its own model, Clef, the same day as Amazon.
What makes this interesting is that three established companies raced to commoditize, for free, the exact niche a small startup had just carved out. TypeSafe barely had an hour before the bigger players caught up. Jev reportedly ran a task for $2.94 that would cost $372 on a frontier model, which is the entire reason this category was invented, and why big players moved this fast to give it away.
Oura’s IPO delay signals a broader market stall
Oura pulled its Nasdaq IPO one day before pricing, even though demand was reportedly four times higher than the shares on offer. The company blamed uncertainty in the IPO market, but according to a source, the real issue is that shares were set to price at the low end of its $40-to-$44 range.
Oura isn’t struggling. It’s profitable, expects 90% revenue growth this year, and has 5.7 million paid members. And yet, when a healthy, oversubscribed company still backs out instead of pricing low, something bigger than that company is going on.
Rising bond yields and high oil prices get most of the blame. Two other companies delayed IPOs this month too. Anthropic is reportedly doing the same, likely until after the midterms. If Anthropic, with the strongest IPO story in tech this year, is also waiting this out, it indicates that even the safest bet in the market isn’t safe enough to price right now.
Google launched Gemini 4 Argon, its most powerful model to date
Google introduced Gemini 4 Argon, its most powerful model, built for long, complex tasks in coding, enterprise work and cybersecurity. Introductory pricing is $2 per million input tokens and $10 per million output tokens. It undercuts GPT-6 Astra’s $10/$50, before rising to $4/$20, which matches Claude Opus 5.5. However, independent evaluators found a mixed picture where Argon ties GPT-6 Astra on Artificial Analysis’s overall index, trails Claude Opus 5.5, and ranks eighth on Arena’s Agent leaderboard.
Even though its benchmarks look strong, almost nobody gets to use it. At launch, Argon is limited to vetted cyber defenders through Google’s Fairwind program, while paid API customers and Ultra subscribers wait with no firm date. Google wants Argon to prove it’s still a frontier AI leader. But almost nobody outside one trusted group can use the model right now. So for most people, this is a marketing claim, not a real product yet.
SpaceX Launches Google AI Chips Into Orbit
SpaceX launched Google’s Tensor Processing Units into orbit aboard Planet Labs satellites as part of a test for space-based data centers. The experiment explores whether AI compute workloads can be run efficiently in orbit to reduce latency.
Anthropic Pushes Opt-Out Model in Australia
Anthropic advocated for an opt-out model for AI training on Australian content, claiming AI could transform the economy. Australian broadcasters ABC and SBS pushed back, warning the technology should face media regulation to prevent news cannibalization.
Armadin Nabs $255M at $2.5B Valuation
Kevin Mandia’s cybersecurity startup Armadin raised $255.5 million at a $2.5 billion valuation in a Series B round co-led by a16z and Accel. The company uses AI agent swarms to detect vulnerabilities and test enterprise security defenses.
Devoted Health Raises $1.2B Funding Round
Medicare Advantage insurer Devoted Health announced nearly $1.2 billion in fresh funding, including $555 million in Series G equity. The company enrolled hundreds of thousands of new senior members this year and is accelerating its growth trajectory.
Flow Engineering Secures $50M for AI Hardware
Flow Engineering raised $50 million in early-stage funding to bring agentic AI to hardware systems design. The platform aims to help engineers develop and integrate complex hardware lifecycles faster by automating parts of the engineering workflow.
CScale Launches with $188M for Data Centers
Optical interconnect startup CScale emerged from stealth with $188 million in funding, about three-quarters from a recent Series C round co-led by Atreides Management. The company develops scale-up interconnect technology for data center infrastructure.
Judge Dismisses Google AI Overview Antitrust Suits
A federal judge dismissed antitrust lawsuits filed by Chegg and Penske Media against Google’s AI Overviews feature. The companies had accused Google of using AI-powered search summaries to siphon traffic away from their websites.
Flytrex Cuts Drone Delivery Costs 60 Percent
Autonomous drone delivery company Flytrex launched rooftop docking stations and predictive AI that cut delivery costs by 60 percent and halved delivery times. The company uses AI modeling to pre-position drones based on anticipated demand patterns.
Inbolt Raises $12.5M for Robot Vision
Inbolt secured $12.5 million in funding led by Shift4Good to bring real-time vision and intelligence to industrial robots. The technology enables robots to see, think, and adapt dynamically to changing production environments without manual reprogramming.
Cloudflare Launches Basin Data Platform
Cloudflare unveiled Basin, a serverless data platform promising reduced egress fees for analytics workloads. The new offering extends Cloudflare’s infrastructure beyond CDN and security into direct competition with cloud data warehouse providers.
Tesla Adds Emergency Drive-Away Feature
Tesla launched an Emergency Drive Away feature that lets drivers shift into gear and leave while still plugged into a Supercharger. The update came two months after a deadly shooting at a Supercharger station where victims were trapped while charging.
FTC to Regulate AI Without New Rules
The Federal Trade Commission announced plans to regulate AI companies including OpenAI and Anthropic using existing consumer protection laws rather than drafting new AI-specific rules. The approach aims to enforce accountability without waiting for congressional action.
OpenAI Fires Three Safety Researchers
OpenAI confirmed it dismissed three safety researchers for allegedly mishandling sensitive company information. According to the Wall Street Journal, the researchers shared confidential material with a third party in violation of company policy.
Prompt: AI ROI Calculator
When to use this?
Use this when you’re evaluating an AI initiative and need to turn productivity gains, costs, and business impact into a clear ROI estimate. It’s especially useful before presenting an AI investment case to finance or leadership.
Act as an AI ROI analyst. Help me estimate the potential ROI of an AI initiative using the information I provide. Ask me for the following inputs, one at a time: AI use case and business process Number of employees affected Average employee cost per hour Current time spent on the process per employee Expected time saved with AI Frequency of the process Expected improvement in revenue, conversion, quality, or other measurable outcomes AI tool, software, implementation, and training costs Any ongoing annual costs Expected adoption rate Expected implementation timeline Then calculate: Annual hours saved Annual labor savings Additional measurable business value Total annual AI costs Net annual benefit ROI percentage Payback period ROI under conservative, expected, and optimistic scenarios Show the assumptions and calculations clearly. Separate hard-dollar savings from estimated or non-financial benefits. Flag any assumptions that could materially change the result, and tell me which 3 inputs I should validate first. Use this formula for ROI: ROI = (Total annual benefit − Total annual AI cost) / Total annual AI cost × 100 Don't invent missing numbers. If I don't know an input, use a clearly labeled assumption and show how the result changes if that assumption is higher or lower.